id
stringlengths 24
24
| idx
int64 0
402
| paragraph
stringlengths 106
17.2k
|
|---|---|---|
648741004f8b1884b731451d | 5 | used GA to design single-stranded DNA grafted colloids . These authors were able to reproduce the experimentally validated phase diagram and additionally identify the formation of four previously unobserved crystal structures. Kim et al. demonstrated one of the first datadriven inverse design methods of new polymers having high band gap and high glass transition temperature that is relevant for high-temperature and high-energy density dielectrics . They successfully identified new polymer structures with the desired properties. |
648741004f8b1884b731451d | 6 | In this work, we follow a similar procedure to Kim et al. to design new polymer membranes with the desired selectivity and permeability for CO 2 separation from N 2 and O 2 . First, we start by assembling a library of gas permeabilities corresponding to the experimental studies of various polymers. Next, we train multiple ML models based on various fingerprints to determine which ML model performs the best in predicting gas permeability. Then, we use our ML models to drive a GA for 100 generations and create more than 16000 new polymer structures. We also use different fitness functions to design the best possible polymers given our initial data set. |
648741004f8b1884b731451d | 7 | Application of this combined ML-GA framework results in the discovery of more than 20 new polymers that are above both the CO 2 /N 2 and CO 2 /O 2 Robeson upper bounds, many of which contain aromatic functional groups along with oxygen-and nitrogen-motifs, aligning with experimental observations that show imines and polyethers as promising polymer membranes for CO 2 separation . |
648741004f8b1884b731451d | 8 | We compiled a literature database of permeability for three gases -CO 2 , N 2 , and O 2 -in a variety of polymers at a temperature range of 300-330 K. The number of data points for each gas is different due to the availability of data in the literature, so we only considered polymers that have permeability measurements for all three gasses. This resulted in 780 different polymers in our library, which represent a sizable portion of the polymers that are typically included in the most up-to-date Robeson plots. The selectivity versus permeability data for CO 2 /N 2 and CO 2 /O 2 are shown in Figure . We see that there are only three polymers that are above the CO 2 /N 2 Robeson upper bound . These polymers have a benzene ring and ether oxygen functional groups in common, which are known to be favorable for CO 2 separation. There are more than ten polymers that are above the CO 2 /O 2 upper bound as shown in Figure . The list of all polymers in our library and permeability measurements are provided in the SI. |
648741004f8b1884b731451d | 9 | The first step in applying ML models to evaluate physical properties is choosing an appropriate mathematical form to be used as input. This is commonly known as featurization (or fingerprinting in the chemo-informatics literature) and it is of critical importance to the quality and interpretability of the ML models. We start with generating the simplified molecular-input line-entry system (SMILES) representations of our polymers based on their repeating units. We cap the two ends of the monomer structure with hydrogen atoms to create a consistent data set. Based on our SMILES strings, we use two common fingerprints in the literature, the Extended Connectivity Fingerprint with bond diameter four Angstroms (ECFP4) and the Molecular ACCess System (MACCS) fingerprint. MACCS is a common substructure keys-based fingerprint consisting of a binary vector of 166 bits depending on the presence of certain substructures or features from a given list of structural keys . ECFP4 is an example of a topological fingerprint that is based on analyzing all the fragments of the molecule by looking at the environment of each atom up to a set radius, and then hashing every one of these environments to create the fingerprint. One needs to be careful when using hashed fingerprints because a bit cannot be traced back to a given feature, and this may result in a given bit corresponding to more than one different feature, which is called "bit collision" . We use ECFP4, based on the Morgan algorithm , which is a 2048 bit fingerprint as implemented in RDKit. Figure shows the comparison for predicting CO 2 permeability with the random forest regression model using both fingerprints. We fit and plot the logarithmic permeability values to better visualize the data set. Both fingerprints result in R 2 value of 0.982 for the training set. However, R 2 for the test set is considerably higher when we use ECFP4 as shown in Figure . |
648741004f8b1884b731451d | 10 | We also compare the root mean square error (RMSE) of the fits for both fingerprints. The test set RMSE with ECFP4 fingerprint is 0.131, and the test set RMSE with MACCS fingerprint is 0.161. Thus, we use ECFP4 to train and test our ML models for the rest of this paper. We start with randomly splitting our data set into one of two categories for each gas; one is used for training the ML model, while the other is initially withheld during training. The training data sets were 80% of our total database for each gas, which represents more than 600 polymers for each gas. We then apply the trained model to the remaining 20% of the polymers (test set) and use these data as verification of the model's accuracy. Then, we employ various ML models on the training sets including support vector regression (SVR), k-nearest neighbors (KNN), decision tree, and random forest regression. Next, we compare the predictive power of these popular ML regression models on CO 2 permeability values. First we study SVR, which has the ability to consider non-linearity in the permeability data . SVR results in R 2 value of 0.84 and 0.203 RMSE on the test set. KNN regression, which predicts the target value by local interpolation of the targets associated with the nearest neighbors in the training set, results in R 2 value of 0.822 and 0.242 RMSE on the test set. Then, we employed a decision tree regression model, which uses a tree structure and inference layer to achieve the final decision of the modeling results . Decision tree regression performs better than both the SVR and KNN regression with R 2 value of 0.881 and 0.148 RMSE on the test set. |
648741004f8b1884b731451d | 11 | Finally, to make a more accurate prediction, we used a random forest regression model, which is an ensemble learning method for regression that combines predictions from multiple decision tree models. As expected, random forest predictions are better than all the other algorithms that we have tried with R 2 value of 0.941 and 0.135 RMSE on the test set. Figure summarizes the different ML regression models that we have tried. We note that Yang et al. compared random forest regression models with deep neural networks (DNN) and showed DNN model performs better than the random forest regression model . However, DNNs typically require much more data than what is available for this study. |
648741004f8b1884b731451d | 12 | To determine where on the Robeson plot a polymer is located, we need to be able to predict the CO 2 /N 2 and CO 2 /O 2 selectivity as well as the CO 2 permeability. The ideal selectivity α i/j for the gas pair is the ratio of the permeabilities P i and P j . Thus, we need ML models to predict N 2 and O 2 permeability as well. Because the random forest regression model is the best performing model for the CO 2 permeability, we have continued using random forest regression for the N 2 and O 2 permeability. Figure shows model predictions for the N 2 and O 2 permeability. |
648741004f8b1884b731451d | 13 | For both gases we can predict the gas permeability with R 2 values higher than 0.9. The RMSE for N 2 and O 2 are 0.171 and 0.147, respectively. This demonstrates that we can predict all three gas permeabilities accurately with the random forest regression model. Now that we have established an accurate ML model to predict a polymer membrane's performance with respect to the Robeson upper bounds, we start designing new polymers with a GA and evaluate their performance on the fly with these ML models. |
648741004f8b1884b731451d | 14 | The first step of the GA is to construct the "gene pool" that will be used to create the initial parent polymers. We used the "Breaking of Retrosynthetically Interesting Chemical Substructures" (BRICS) algorithm as implemented in the RDKit Python package to get the chemical building blocks, or fragments, from our polymer library . A total of 79 unique fragments were extracted from 780 reference polymers. Figure shows six functional groups that appear most frequently in our library. To initiate the GA process, 100 parent polymers consisting of 4 building blocks in their monomer unit were created in the first generation. The fragments were chosen randomly from our gene pool of the 79 chemical fragments. Then, 15 families with the smallest Tanimoto similarity score , with 3 parents in each family, were chosen to perform crossover and mutation operations to alter their sequence of chemical building blocks, resulting in 12 offspring polymers in each family. During crossover, two parents generate an offspring by combining one random segment from a parent with another random segment from the other parent. The segmentation point of a parent polymer was chosen according to a Gaussian distribution with a mean at the center of the sequence and standard deviation of 0.3 blocks. We also applied mutation operations on 60% of the genes to increase the chemical diversity, where we randomly selected the building block in the sequence and replaced it with a new building block randomly chosen from the list of the 79 blocks. In each GA iteration, the top performing offspring polymers with the highest fitness evaluation were retained as parents to create the next generation offspring polymers. We also assigned 10% migration rate between different families, whereby the highest-scoring polymers that were not selected as a parent migrated to a different family. An essential component in this evolutionary process is the polymer property estimation, which traditionally has been evaluated by experiments that are very time-consuming and expensive; here, we use our ML models for on-the-fly polymer property estimation. |
648741004f8b1884b731451d | 15 | We ran the GA for 100 generations and generated more than 16000 new polymer structures as shown in Figure . All the new polymers generated with the GA are reported in the SI, where we highlight the top performing 100 polymers. We optimized multiple parameters in the GA framework to guide the evolutionary process towards the targeted design area. First, Figure : The six most chemical functional groups that appear in our library of experimental polymers using the BRICS algorithm. "A" represents the binding sites. |
648741004f8b1884b731451d | 16 | we ran the GA with an ML model trained with two different fingerprints, ECFP and MACCS keys, and found that the fingerprint used in the ML does not influence the top performing polymers identified from the GA framework. Next, we tried running the GA for an additional 100 generations to see if running the GA for longer will result in better performing polymers, but found that the additional iterations did not result in any improved polymer structures. Finally, we tried multiple fitness functions to optimize the evolutionary trajectory. To optimize both CO 2 /N 2 and CO 2 /O 2 selectivity as well as CO 2 permeability, one needs a fitness function that includes all three metrics. However, the functional form of the fitness function is not clear a priori. We tried the fitness function log (P CO 2 ) × α CO 2 /N 2 × α CO 2 /O 2 and showed that GAs using this fitness function failed to identify new polymer structures that are above both upper bounds. We found that the fitness function based on the actual P CO 2 permeability values |
648741004f8b1884b731451d | 17 | result in several polymers that are above both the CO 2 /N 2 and CO 2 /O 2 upper bounds as shown in Figure . Because CO 2 permeability values are generally orders of magnitude larger than the selectivity values, this model favors the polymers that have higher permeability, thus biasing the GA towards better performing polymers. |
648741004f8b1884b731451d | 18 | We used the BRICS algorithm on the GA-generated polymers to understand which chemical building blocks are frequently observed in the top performing polymers. The top six frequently observed functional groups with the 100 fittest GA-generated polymers are shown in Figure . We identified a total of 464 chemical fragments within the fittest 100 polymers, and 18% of the fragments were pyridine functional groups. More than 70 polymers in the top performing GA-generated polymers have the pyridine functional group in their repeating unit. This is by far the most frequently observed functional group, which is then followed by benzoxazole with 3%. We observe benzoxazole functional group in 13 polymers within the 100 fittest GA-generated polymers. Similarly, benzene, phosphonamidic acid, naphthalene, and dibromobenzene functional groups are also observed in the top performing polymers generated with the GA. We show six example polymer structures that have high fitness function values in Fig- . We note that these polymers have pyridine, benzoxazole, benzene, and phosphonamidic acid functional groups, which we identified as the most abundant functional groups with the BRICS algorithm. These polymers also include oxygen-, sulfur-, and nitrogen-containing motifs, similar to the three experimental polymers that are above the CO 2 /N 2 upper bound. |
648741004f8b1884b731451d | 19 | Oxygen-and nitrogen-containing motifs are reminiscent of imines and polyethers, which are known to be high performing polymer membranes. Interestingly, our initial analysis using the Polymer Genome software suggests that most of the top 100 fittest polymers have high glass transition temperature-well above the standard operating conditions (> 400 K). We note that the three polymers in the experimental data set that are above the CO 2 /N 2 upper bound also have glass transition temperatures around 400 K. We speculate that the superior performance of these glassy amorphous polymers for gas separation may be due to their high fractional free volume and high number of microvoids . It remains an open question whether or not these polymers are easily synthesizable and easy to implement as membranes given their complicated chemistry. Further computational and experimental studies will be required to better understand these polymers and their efficacy as membrane materials. |
648741004f8b1884b731451d | 20 | We constructed an ML-driven GA to tackle the inverse design problem of polymer membranes for CO 2 separation. We showed that the hashed-based ECFP4 yields lower predictive errors on the test sets than the substructure keys-based fingerprints. We presented different regressionbased ML models, where random forest regression models resulted in the lowest RMSE and highest R 2 values for both the test and the training set. Although random forest regression models can successfully predict both the gas permeability and the selectivity, we used models trained on the gas permeability, since these models have better predictive power than the models trained on the selectivity. After obtaining the ML models to predict the performance of any polymer membrane candidate, we implemented a data-driven inverse design algorithm to efficiently explore the polymer material space. In theory, one can use any inverse design algorithm for such a problem, but we have implemented a GA since it has been successfully used for other polymer applications. We created the gene pool using the BRICS algorithm on the experimental data and obtained 79 unique genes to initiate the GA process with 100 parent polymers that have 4 genes in their monomer unit. Fitness function is a key driving parameter in the GA, and as such, we used |
648741004f8b1884b731451d | 21 | determine the fitness of the polymers. We performed crossover and mutation functions for 100 generations to create more than 16000 polymers during the GA process. Among these 16000 polymers, we were able to identify more than 20 new polymers that are above both CO 2 /N 2 and CO 2 /O 2 upper bounds. |
648741004f8b1884b731451d | 22 | Three key points emerged from our analysis on comparing popular fingerprints. First, hashed-based fingerprints result in lower predictive errors on the test sets than the substructure keys-based fingerprints. However, hashed-based fingerprints have one main disadvantage compared to the substructure keys-based fingerprints, which is not having a one-to-one correspondence between the fingerprint vector and the chemical structure. This is not the case with the substructure keys-based fingerprints, where each bit corresponds to a predetermined substructure. This disadvantage does not affect our framework since we do not go back-and-forth between the fingerprint and the chemical structure, and only use the fingerprinting when evaluating the fitness function in the GA framework. Not performing crossover and mutation functions on the fingerprints makes it possible to overcome this main disadvantage associated with hashed-based fingerprints. Next, we note that the top performing polymer candidates identified within this framework does not depend on which fingerprint is used in the GA. The relative strength of the polymers with each other are similar with the two different fingerprints. The main difference with using different fingerprints is the absolute value of the fitness function, since the ML models trained on the substructure keys-based fingerprints tend to underestimate the gas permeability. Finally, using the proper descriptor for a given application is still an open question in the polymer informatics field, but we have shown that most often it does not affect the final result. However, we emphasize that more sophisticated descriptors, like physical descriptor vectors that include bulk properties of the polymer membranes, may make a difference in the top performing polymers identified from the GA . For example, we believe the glass transition temperature is a key property for polymer membranes, and including this information in the fingerprint can lead to better performing polymer membranes within our framework. |
648741004f8b1884b731451d | 23 | The main bottleneck in switching from popular fingerprints to physical descriptors is gathering consistent measurement data from hundreds of different experimental papers. The only way to overcome this bottleneck is to create our own data sets using computational simulations so that we can consistently calculate the physical properties of interest for each polymer. |
648741004f8b1884b731451d | 24 | We have demonstrated that a random forest regression model performs best when predicting the gas permeability and selectivity of polymer membranes. Because there is a significant difference in the R 2 values and the RMSEs, we use random forest regression models in the entire framework. However, random forest algorithms, like essentially all data-driven methods, are intrinsically interpolative . They are only suited to optimize properties within the bounds of the data the model was trained on. Models can still generalize, and interpolate "between molecules" in some abstract design space ( ), but they will not make accurate predictions outside of this space. Thus, the performance of the new polymer structures identified in this framework depends on the initial data set and the range of selectivity and permeability values covered. One way to address this challenge is by using computer simulations combined with e.g. an active learning loop to curate a polymer library will enable us to expand the number of data points included in the ML framework. It is important to note that regardless of the datadriven approach, ML models will always have a strong dependency on the initial data that they are trained on. The only way to surpass this main limitation is to move towards active learning algorithms where we give the algorithm the ability to "learn" and draw inferences from its experience to accelerate the evolutionary process . New molecules generated as part of the GA procedure could be screened by an uncertainty-quantifying algorithm, and when confidence is low, new simulations can be run to acquire new ground truth data, which can then be used to retrain the model. This is only possible if we use computer simulations (or a very highthroughput, autonomous experiment) to curate the data since we will need to make on-the-fly property estimations with an active learning framework. |
648741004f8b1884b731451d | 25 | Finally, we identify two significant points for coupling GAs with ML models to design new polymer membranes. First, and most importantly, the fitness function drives the evolutionary process of the GA. It is therefore of utmost importance to select the correct fitness function to direct the GA towards the targeted design area. It is customary to train ML models on the logarithmic permeability values since it narrows the range of the data, hence resulting in more accurate models. Thus we tried the fitness function log (P CO 2 ) × α CO 2 /N 2 × α CO 2 /O 2 , aiming to maximize both the gas permeability and selectivity throughout the evolutionary process. However, this fitness function was not able to push the GA towards the targeted design area. Even though ML models trained on logarithmic permeability have slightly higher predictive power, the small numerical value of the logarithmic permeability diminishes the importance of the permeability contribution to the fitness function. On the other hand, with a fitness function that includes the absolute value of the permeability P CO 2 × α CO 2 /N 2 × α CO 2 /O 2 , we were able to push the evolutionary process toward the targeted design area and identified more than 20 new polymers that are above both CO 2 /N 2 and CO 2 /O 2 upper bounds. We attribute the superior performance of this fitness function to the fact that the absolute value of the permeability is usually two orders of magnitude higher than selectivity values. This analysis also shows improving the selectivity with the GA is much harder than the permeability since the fitness function becomes insensitive to the selectivity values when we include the gas permeability. In the future this can be avoided by normalizing the parameters where the normalization would negate this effect. Next we emphasize that, our GA was able to converge within 100 generations, since running the algorithm for an additional 100 generations did not result in any superior polymer membranes. With 100 generations and 4 initial building blocks in the first generation, a total of 17571 new unique polymer structures were created. With 79 unique genes, the number of sequences that can be generated by the GA is at least 79 4 (since longer sequences are generated throughout the evolutionary process). This suggests that the GA converges very fast, only exploring less than 1% of the possible polymer material space. We decided to use 4 genes with the initial generation because we have a relatively small gene pool. Using more building blocks with the initial generation could have created more complicated structures throughout the evolutionary process. This can be further explored when we have a larger gene pool. |
648741004f8b1884b731451d | 26 | Our approach demonstrates successful implementation of an ML-driven GA to design polymer membranes for CO 2 separation, but more importantly, this framework can be used to design polymer structure for any application (e.g. ion separation membranes and polymer electrolytes for batteries), where there is a constrained optimization problem. The main limitation of the current framework arises from its dependence on the initial experimental data. Curating the data with computer simulations is a possible way to overcome this limitation. With better control over the initial data set we will be in a position to explore more sophisticated descriptors and switch to an active learning framework where we make on-the-fly property estimations. |
61366473fc08e36aaf8c3cf6 | 0 | The development of dynamic covalent chemistry (DCC) sensors can fundamentally impact the disparate fields of biology, medicine, and materials science. At the heart of successful DCC-based sensors is a robust molecular recognition strategy that involves the binding of and the chemical or conformational response to the analyte. Recognition of molecule type is often comparatively simple to engineer since it relies on predicable functional group-specific reactions (e.g., nucleophilic attack of an alcohol on an iminium). In contrast, a difficulty of sensor design often stems from having to discriminate between different molecules bearing the same or similar functional groups or having to respond to specific analyte stereochemistry. Solving these types of challenges often involves creating sensors that recruit the subtle additive effects of modest (<2 kcal/mol) noncovalent interactions (NCIs). However, the development and understanding of these types of sensors is hampered by the difficulty of deconvoluting individual effects of the multiple, competing NCIs involved. |
61366473fc08e36aaf8c3cf6 | 1 | As examples, the challenge of designing sensors based on dynamic covalent chemistry coupled with weak NCIs is evident in the field of chiroptical sensing (Figure ). Many chiroptical sensors 1 feature a configurationally labile set of CD-active chromophores that interact via exciton coupling (e.g., + and -configurations, depicted in red and blue, respectively). In the absence of a chiral analyte, the (+)-1 and (-)-1 configurations are enantiomeric and exist in a 1:1 equilibrium ratio. As a result, the Cotton effects from each enantiomer of the sensors cancel with no net CD signal observed. However, upon binding an enantioenriched analyte, the NCIs between the analyte 2 stereocenter and the optically active chromophores (depicted with a grey arrow) can lead to a net CD signal, which in turn is used to determine the enantioenrichment of 2. When the stereocenter is adjacent to the binding functional group (2, n = 0), repulsive steric interactions influence the relative twist of the chromophores, often leading to a substantial CD-signal (e.g., > 10 mdeg). When the stereocenter is remote (2, n ≥1) the steric interactions necessarily diminish, and multiple weak NCIs (including attractive interactions) could presumably induce a measurable CD-signal. However, there are a dearth of reports for chiroptical sensing of remote stereocenters. 6 One of our groups has been particularly interested in developing chiroptical sensors of enantioenriched alcohols (Figure ). We have shown that 2-formylpyridine (4) can efficiently react with dipicolylamine (5) and Zn(OTf)2 in the presence of an acid catalyst (chloroethylmorpholine hydrochloride, CEM•HCl) to form hemiaminal 6 (Figure ). The pyridine rings in 6 adopt enantiomeric CD-active helical configurations (P and M, depicted in red and blue, respectively). Upon binding of a chiral alcohol 7, a series of diastereomers form by virtue of the pyridine helical twist, the chiral alcohol, and the newly formed point stereocenter of the hemiaminal ether functional group. If the alcohol is enantioenriched, one configuration at the hemiaminal ether group dominates, which induces a bias toward a P or M twist of the pyridines, leading to exciton coupling and thus an observable CD signal. The magnitude of the CD signal can be correlated through a linear free energy relationship to differences in steric size of the R-groups on the stereocenter of the alcohol. This sensor assembly has proven to be robust in detecting the enantioenrichment of a-stereogenic alcohols in many settings. More recently, we developed complex 8 that is capable of sensing the more challenging b-stereogenic alcohols (Figure ). The design strategy hinged on the incorporation of an "appendage" off the pyridine ring closest to the hemiaminal group that could interact with the more remote stereocenter. While the assembly formation is identical to 9, the chromophore responsible for sensing the chirality of the alcohols is an atropisomeric biaryl motif (depicted in red and blue). Importantly for the present work, however, designing sensors for the more challenging g-stereogenic alcohols 11 remains a tremendous challenge (Figure ). Merging the need to develop sensors for g-stereogenic alcohols with the challenges of identifying what structural features are required to achieve this goal, we sought to develop an integrated synthetic and data science-based workflow (Figure ). With such a remote stereocenter, significant steric interactions with the three pyridines or the atropisomeric chromophore were anticipated to be minimal, while the possibility of attractive NCIs could arise that would place a twist into either chromophore. Ideally, this workflow would allow one to simultaneously optimize, as well as understand, the performance of chiroptical sensors that require multiple weak NCIs to function effectively. Four key stages were envisioned: 1) a simple machine-learning (ML) enabled strategy would be used to carefully identify highly diverse appendages (blue in structure 10) that are synthetically accessible, 2) the selected compounds would then be synthesized and tested, 3) the resultant data would be used to train statistical models that correlate structural features to CD signal intensity, and 4) interrogation of the models would provide insight to inform next generation designs. This workflow is iterative in nature, wherein the mathematical model is applied to quantitatively predict the performance of new assemblies and ultimately deconvolute the noncovalent interactions critical to the success or failure of the chiroptical sensor. Herein, we present the successful application of this workflow to the identification and interrogation of chiroptical sensors of g-stereogenic alcohols. Further, we were able to bolster the insight acquired from statistical modelling with DFT-level calculations to probe the underlying NCIs responsible for the observed sensor performance. |
61366473fc08e36aaf8c3cf6 | 2 | To effectively use this workflow to design chiroptical sensors for gstereogenic alcohols, we had to select which modular structural element of the assembly to systematically vary (Figure ). We identified the A-ring (highlighted in blue) as ideal. Diverse moieties could be easily introduced using a metal-catalyzed cross-coupling reaction and would also presumably be well positioned to interact with the g-stereocenter. |
61366473fc08e36aaf8c3cf6 | 3 | To ensure that the biaryl aldehydes 10 selected for evaluation are chemically diverse, we employed a data-driven substrate selection strategy (Figure ). This was comprised of an initial generation of a data-rich virtual library of arylboronic acids 13 followed by a machine learningbased selection strategy. The virtual library was first curated by computationally assessing all commercially available and in-house aryl boronic acids (a total of 5,136 boronic acids) (Step 1). To accomplish this, the geometry of each substrate was minimized and then 1,357 computationally inexpensive 2D and 3D Quantitative Structure Activity Relationship (QSAR) parameters -simple molecular descriptors routinely used in medicinal chemistry -were calculated for each boronic acid (Step 2, see SI for details). With this computational data set, the boronic acids were analyzed based on their similarity in descriptor space by performing principal component analysis (PCA) to reduce the dimensionality of the computed descriptors. A small, representative set was then selected for biaryl aldehyde synthesis by employing K-means clustering. This machine learning algorithm is used to group or "cluster" structures based on structural similarity (Step 3). Through this process, nine boronic acids (described below) were identified from distinct clusters, procured, and subjected to a Suzuki cross-coupling with aryl bromide 14 to produce the initial library of biaryl aldehydes 10a -10k (Step 4). We note that the selection strategy depicted in Figure resulted in a small, highly diverse library of candidates. Intuitively, this is inferred from a simple inspection of the structures and substitution patterns of 10a -10k. The library features both electron rich and electron poor arenes (e.g., 10a and 10d, respectively), substrates with extended p surface area (10f, 10g, and 10h), and a multitude of substitution patterns (e.g., 10c, 10e, and 10k). Moreover, the chemical diversity resulting from our workflow can be visualized (each member plotted as a blue point) using a Uniform Manifold Approximation and Projection (UMAP) visualization plot (see plot depicted in step 3). With UMAP, similar boronic acids are positioned close to one another in their descriptor space as demonstrated by their grouping in "islands". The compounds that were selected from our workflow (plotted with yellow circles) are well spread across the representation, reflecting the statistical diversity of our library. It should be noted that the UMAP axes enable visualization of the variance in the two plotted dimensions, similar to a principal component representation. We next turned our attention to the evaluation of assemblies 12, which were readily formed from the reaction of biaryl aldehydes 10 with Zn(OTf)2, dipicolylamine, and an enantioenriched g-stereogenic alcohol (either 11a or 11b, Figure ). The resultant assemblies 12a and 12b were assessed via CD spectroscopy (subscript a and b denote the use of alcohols 11a and 11b, respectively). In nearly all cases, we observed a common Cotton effect at 267 nm (see representative spectra, Figure ). On the basis of previous reports, it was concluded that this feature arose from exciton coupled circular dichroism (ECCD) of the helical pyridine chromophore (depicted in blue). The intensity of the CD signal varied substantially depending on the structure of 12, with signals of up to 22.7 mdeg observed for 12da and near-zero signals for assemblies 12bb, 12gb, and 12hb (Figure ). The diversity of responses is likely a result of the intentional incorporation of disparate In order to correlate CD intensity to the structural features of 12, we next employed a two-step statistical modelling protocol (Figure & B). This was accomplished by computing a range of physical organic descriptors to reflect the physical properties of 12a -12k. These were computed using a simplified chemical structure 15, which captured the essential structural elements anticipated to influence the CD response to the chiral alcohols. The descriptors included sterimol values (steric measurements), global electronic terms such as HOMO/LUMO energies and polarizability, as well as local electronic terms reflected by NBO charges on various atoms (see SI for full parameter list). Using a forward stepwise linear regression algorithm to correlate experimental CD intensities to computed parameters, a fourterm statistical model was found (Figure ). This model included one parameter to account for the alcohol 11a/11b and three derived from the biaryl pyridine motif. The alcohol parameter (classifier) was a simple unitless value (1 for 11a, -1 for 11b), used to assign which alcohol was used in the assembly. The remaining three parameters, NBOipso, DEFMO, and ortho B5 (a sterimol value), were derived from 15 (Figure ). While the parameters NBOipso and DEFMO had a relatively even distribution of values across 15a -15k (see SI), the B5 (Sterimol parameter reflecting size) for the ortho substituent had only three discrete values. Assemblies stemming from biaryl aldehydes with H at both A-ring ortho positions (12a -12i) all had an ortho B5 of 1.1Å and only 12j and 12k had larger measurements. Therefore, the parameter ortho B5 can be considered a classification term that reflects whether 12 has an ortho substituent. |
61366473fc08e36aaf8c3cf6 | 4 | The first iteration of the workflow accomplished one of our two initial goals, i.e., it identified which biaryl aldehyde (10j) would lead to assemblies with large enough CD signals to accurately determine the ee of both 11a and 11b. To our knowledge, this is the first chiroptical sensor that can assess a stereocenter so remote to the binding functional group. However, the second primary goal, to deconvolute the NCIs that underpin a successful sensor, was not yet realized. The model depicted in Figure is statistically sound, but the descriptor terms do not give clear insight into the NCIs that determine the magnitude of the CD signal. For instance, it is challenging to understand the physical significance of the NBO charge on the ipso carbon (NBOipso, highlighted with a green sphere). Furthermore, the need to classify two aldehydes that incorporate an ortho-substituted A-ring (only 10j and 10k) reduces the utility of the statistical modelling approach. |
61366473fc08e36aaf8c3cf6 | 5 | As a result, we revisited the workflow discussed above for a second iteration with the goal of exploring regions of chemical space that should be further sampled. Specifically, the under representation of ortho substituents in the initial library led us to the hypothesis that incorporation of additional examples would enhance our application of regression analysis. In turn, the improved regression model would likely provide a more detailed understanding of the origin of the NCIs at play in determining large CD signals. |
61366473fc08e36aaf8c3cf6 | 6 | To better understand of the role of ortho substituents in the chiroptical sensing of g-stereogenic alcohols, we next focused on the selection, synthesis, and evaluation of aldehydes 10l -10p (Figure ). Boronic acid precursors 13l -13p were selected using the ML-based selection workflow described in Figure (13l -13p depicted with orange diamonds, Figure ). It should be noted that this selection protocol focused on a narrower region of chemical space (ortho-functionalized compounds only) and that this is reflected by the close grouping of 13l and 13n -13p on the UMAP plot. The resultant assemblies 12l -12p were then evaluated via CD spectroscopy. As was anticipated from the MLR model shown above, ortho-functionalized assemblies 12j -12p tended to give comparatively higher CD intensities than 12a -12i (Figure ). Nevertheless, the varied substituents in the aldehyde structures provided meaningful trends difficult to ascertain by simple inspection. Once more, we turned to statistical modeling to assess assemblies 12a -12p (Figure ). The best model obtained was able to adequately account for both assemblies 12a and 12b suggesting common NCIs intrinsic to the biaryl motif are similarly important whether alcohol 11a or 11b are used. The model featured a good correlation (R 2 = 0.91) and robust internal validation measures (Q 2 = 0.85 and 4-fold = 0.84). Furthermore, an external validation was prepared by initially partitioning the data into training and test set (80/20 split, see SI for details) and then evaluating the ability of the model to predict the test set data points. It was found that the model reliably predicted the CD signals of the test set (test R 2 = 0.91). |
61366473fc08e36aaf8c3cf6 | 7 | As in the previous iteration of MLR modeling, we evaluated a multitude of steric and electronic parameters in this second model (see SI for more details). Those parameters that were included in the final model are depicted in Figure & 8C and include the same alcohol classifier and DEFMO terms used in the previous model, as well as two new steric terms, Bmax 2Å and Bmax 8.5Å. The latter two parameters are advanced sterimol descriptors recently introduced by Paton and coworkers. The parameters were measured by first defining an axis along the pyridine C4-C1' bond (depicted with a grey line, Figure ). The maximum and minimum steric measurements perpendicular to this axis (Bmax and Bmin, respectively) were collected at 0.5 Å intervals along the axis. The Bmax measurements collected at 2.0 Å and 8.5 Å led to the best MLR model. In order to further validate the robustness of the statistical model, we synthesized and tested an additional validation set of 4-biaryl aldehydes 10q -10t that were not used for statistical modeling (Figure ). These were synthesized from boronic acids that were already on hand in our laboratory via a Suzuki cross-coupling reaction. The aldehydes were then tested with alcohols 11a and 11b and the CD spectra of the resultant assemblies were measured. The original model was then retrained using all assemblies 12a -12p. We were pleased to find that assemblies 12q -12t were well predicted by the model (validation R 2 = 0.81). It should be noted that the compounds in the validation set featured both electron-poor (12t), electron-rich (12s), and p-extended (12q) arenes and a common 3,5-disubstitution pattern that was absent from the aldehydes used for model generation. This highlights the ability of the model to effectively predict the performance of unique, out-of-sample substrates. |
61366473fc08e36aaf8c3cf6 | 8 | While the second model was only modestly better than the first in terms of regression statistics (R 2 , test R 2 , etc.), it was significantly more interpretable. Both the "hardness" term (DEFMO) and alcohol classifier term were conserved across both models. However, the relatively opaque NBOipso value and the ortho-substituent classifier were replaced by two intriguing steric parameters presented above. These were particularly interesting because of their contrasting correlation to CD signal. While steric bulk was positively correlated to CD signal at 2.0 Å, it was negatively correlated at 8.5 Å. As discussed below, these pronounced distance-resolved effects provided insight into subtle intramolecular NCIs within assemblies 12 that critically effect CD signal intensity. |
61366473fc08e36aaf8c3cf6 | 9 | Armed with a statistical model composed of mechanistically interpretable parameters, we sought to answer the following question: what is the mechanism for the transfer of stereochemistry from the CD-silent point stereocenter of the alcohol to the CD-active terpyridine helix? As stereoenrichment of the terpyridine is the basis for the CD signal, an underlying understanding of this phenomena would provide a blueprint for what structural features make an effective sensor. |
61366473fc08e36aaf8c3cf6 | 10 | A key consideration relevant to this question was whether the configuration at the aminal stereocenter (R vs S) played a role in biasing the terpyridine configuration (M vs P). One of our groups had previously discovered a correlation (R 2 = 0.97) between the dr of 8 and the measured CD signal intensity (Figure ). On the basis of this striking correlation, it was concluded that a point-to-point-to-helix mechanism of stereochemical transfer was operative (Figure ). This can be conceptually deconstructed into two steps: (1) the alcohol stereocenter (highlighted with a blue star) causes one of the hemiaminal ether epimers (stereocenter highlighted with a grey sphere) to be enriched. In turn, the configuration of the aminal influences the energetics to favor one of the two terpyridine configurational twists. It was determined that the (S)-aminal led to (P)-helicity and the (R)-aminal led to (M)-helicity. Therefore, the ratio of aminal epimers directly controls the magnitude of the CD signal. Given the good correlation previously observed between CD signal intensity and dr for assembly 8, we questioned whether a similar relationship existed for the g-stereogenic alcohol-derived assemblies 12 (Figure ). In order to test this, we measured the dr of several assemblies 12a via 1 H-NMR spectroscopy as previous reported. 39,40 Although a similar range of values was observed (dr =1.1 -2.2), there was no correlation with CD signal intensity. This suggested that either: (1) another variable was at play that obscured the correlation, or (2) the hemiaminal ether stereochemistry did not strongly influence the pyridine configuration and, instead, another mechanism for stereochemical transfer was involved. |
61366473fc08e36aaf8c3cf6 | 11 | In order to interrogate the mode of stereochemical transfer from point to helical chirality, we performed a computational study of assemblies 12ba and 12oa, as well as the previous assembly 8 (Figure ). Given the presumed importance of both proximal and distal steric effects revealed by MLR modelling, 12oa was selected because it had an ortho substituent on the A-ring, while 12ba was para-substituted. Additionally, 12oa and 12ba were the best and worst-performing assemblies tested during model construction respectively and would, therefore, reflect the limiting cases. Assembly 8 was selected as it would provide a benchmark into the origin of the impact of hemiaminal-ether stereo-chemistry on the CD response. For each assembly analyzed, both aminal epimers were subjected to a gas-phase molecular mechanics-based conformational search using the OPLS3e forcefield. From this, five distinct conformers were identified for 8 while the conformationally flexible 12ba and 12oa resulted in numerous conformers (92 -145 per epimer). To limit computational resources, the latter two were clustered (17 -32 clusters per epimer) based on atomic RMSD and one representative conformer was selected from each cluster. All conformers from 8 and the selected conformers from 12ba and 12oa were then refined via DFT calculations. Geometry optimization was conducted using the B3LYP functional with the 6-31G(d,p) basis set in Gauss-ian16. Single-point corrections were then carried out with the M06-2X functional and def2-TZVP basis set. Solvation effects of acetonitrile were considered using the PCM solvation model. The lowest-energy structures of both (P, S)-8 and (M, S)-8 (Figure ) were analyzed revealing a large energy difference between the (M) and (P) configurations (DGM/P = 2.2 kcal/mol). This likely stems from the difference in orientation of the alkoxy group between the (P) and (M) conformers. (P, S)-8 positions the C-O bond of the alkoxy group (highlighted in blue) downward with respect to the N-Zn bond (highlighted in red). This positioning of the alkoxy group is termed axial and likely benefits from minimized steric interactions with the pyridine motifs. In contrast, the alkoxy group is projected roughly perpendicular to the Zn-N bond in the (M, S)-8 conformer. This placement of the alkoxy group, termed equatorial, is likely disfavored due to energetically unfavorable repulsive interaction with the pyridine rings. The substantial energetic preference for one terpyridine configuration is consistent with the notion that the hemiaminal ether stereocenter controls the helicity. We next considered the computed structures for 12ba, the worst-performing assembly tested during model generation (Figure ). While we investigated both hemiaminal ether epimers, we will discuss (R)-12ba as it was calculated to be more stable than (S)-12ba (not shown) by ca. 1.8 kcal/mol. It was noted that the 30 conformers assessed at the DFT level of theory could all be categorized into two distinct conformational ensembles: one with (M) and one with (P) terpyridine configurations. We then compared the energies and structures of the lowestenergy conformer from each ensemble. As was the case for 8, the configuration of the terpyridine dramatically impacted the orientation of the alkoxy group for (R)-12ba. In this case, the (P) and (M) configurations coincide with the alkoxy group being placed equatorial and axial, respectively. Unlike 8, however, the energy difference between (P, R)-12a and (M, R)-12a were comparatively small (DGM/P = 0.1 kcal/mol). The low energy difference between helical conformers is consistent with the modest CD-signal observed experimentally. Finally, we turned our attention to 12oa, the best-performing assembly assessed in this study (Figure ). Once more, we will only discuss the lowest-energy hemiaminal ether epimer, which was (S)-12oa. All conformers could again be categorized into two distinct conformational ensembles, each with an opposite pyridine helicity. The (M)-terpyridine configuration resulted from the equatorial alkoxy orientation while the (P)-configuration favored the axial alkoxy positioning. Unlike both 8 and 12ba, in which the axial alkoxy orientation was favored, the conformer with the equatorial orientation of the alkoxy group (M, S)-12oa was preferred, now by 2.7 kcal/mol. This comparatively large energy difference likely results in a substantial enrichment of the (M) terpyridine configuration and is consistent with the large CD-signal observed for 12oa. Our combined computational and statistical modeling efforts suggest that attractive p-p interactions as well as London dispersion forces, play a key role in controlling CD-signal intensity by modulating DGM/P (Figure ). For assembly (R)-12ba, we observed attractive NCIs that stabilize both the (P)-equatorial and (M)-axial conformations. Londondispersion contacts can be seen in the former between the g-phenyl ring and both the pyridine rings and the A-ring ortho-protons (see NCI plots, Figure ). 50,51 However, long-range dispersive interactions between the alkoxy chain and the A-ring para-substituent also stabilize the (M)axial conformation. The presence of distal dispersive contacts (i.e., the 4-bocamino group), therefore, stabilizes (M,R)-12ba relative to (P,R)-12ba and attenuates DGM/P. In contrast, more significant attractive NCIs were observed for the (M)-equatorial conformer of (S)-12oa than in the (P)-axial conformer (Figure ). In addition to extensive dispersion interactions, the former benefits from a T-stacked interaction between the A-ring ring ortho-proton and the g-phenyl moiety (see dark blue attractive T-stacking interaction in the inset, Figure16C). The limited rotation enforced by the ortho isopropyl group results in an increased dihedral angle of the biaryl which, in turn, enhances this Tstacked interaction. |
61366473fc08e36aaf8c3cf6 | 12 | The computational analysis discussed herein, details the complexities of dynamic covalent sensors that exist as a mixture of multiple equilibrating diastereomeric and conformational isomers (i.e. 16, Scheme 16A). While we have analyzed this in detail in the preceding paragraphs, the overarching conclusion can be summarized as follows. Substitution at the A-ring ortho and para positions has opposite, competing effects on the CD-signal intensity. Ortho steric bulk energetically favors one of the helical diastereomers by reinforcing a key T-stacking interaction (i.e., Figure , inset). This leads to a large, net CD-signal from exciton coupling of the energetically preferred helical chromophore. Para steric bulk, in contrast, stabilizes the minor isomer leading to a smaller energetic preference for one terpyridine helix. This means that the exciton coupling of the terpyridine chromophores with opposite helicity will largely offset one another and only a weak net CD signal will be observed. The competing roles of para and ortho substituents are also reflected in the final MLR model (Figure ). Steric bulk at the para position is reflected by the term Bmax 8.5Å and is negatively correlated to CD signal intensity, while steric bulk at the ortho position, reflected by Bmax 2Å, is positively correlated. Taken together this illuminates the following design principle. Installation of an ortho-functionalized A-ring and omission of para-substituents will maximize efficient stereochemical transfer from the hemiaminal ether stereocenter to the terpyridine helix and will thereby maximize the CD signal for chiral alcohols of this type. |
61366473fc08e36aaf8c3cf6 | 13 | The development of DCC-based sensors remains a broadly applicable endeavor within the field of supramolecular chemistry. However, the identification and deconvolution of subtle NCIs that are important to sensor performance remains an unsolved problem. In this study, we used a combination of computational parameterization, statistical modeling, and high-level DFT calculations to develop and gain a detailed understanding of the first reported chiroptical sensor for g-stereogenic alcohols. By performing two iterations of a data-driven optimization workflow, we were able to identify a highly effective sensor and produce a robust, interpretable statistical model. This provided the basis to deconvolute the roles of distal and proximal steric bulk on sensor performance by mathematically relating the distance resolved sterimol parameters Bmax 2 Å and Bmax 8.5 Å to CD-signal intensity. We then performed high-level DFT calculations to understand the physical significance of the steric parameters. These calculations both revealed a likely sensing mechanism and suggested that both distal and proximal substituents influence sensor performance through attractive NCIs. This work demonstrates the effectiveness of an iterative data-driven approach to sensor design and showcases the utility of computational parameterization and statistical modeling to deconvolute competing weak NCIs. We anticipate that the workflows described herein can be readily adopted for the design of other DCC-based assemblies and in the design and understanding of new supramolecular systems at large. It should be noted that 13i and 13j were not chosen using the workflow outlined in Figure . For further discussion see the SI. McInnes, L.; Healy, J.; Saul, N.; Großberger, L. UMAP: Uniform Manifold Approximation and Projection. J. Open Source Softw. 2018, 3, 861. It should be noted that both enantiomers of alcohol 11b were tested in several assemblies. As expected, equal and opposite Cotton effects were observed (see SI for more details). This Cotton effect lmax deviated by ± 2 nm at the most, however, it typically only deviated by ± 0.5 nm. All reported CD-signals were normalized by dividing the measured CD-signal (mdeg) by the absorbance. Our group has extensively employed DFT-parameterization followed by multivariate linear regression for reaction development. For further discussion see references 28-30. It should be noted that assembly 12ca was an outlier and was not included in model development. It is likely that the nitroaromatic moiety engages in distinct p-p NCIs with the g-phenyl ring of 11a that are not accounted for by our models. For further information about parameter collection, see discussion in the SI. As in the modeling depicted in Figure , assembly 12ca was an outlier and was excluded from model development. Luchini, G.; Paton, R. S. DBSTEP: DFT Based Steric Parameters. 2021, DOI: 10.5281/zenodo.4702097. It should be noted that Paton's advanced distance-resolved sterimol descriptors were critically important as the simple sterimol measurements of the arene as a whole were poorly correlated to CD signal. The opposite correlations for proximal and distal encumbrance likely counteract one another when distance-resolved parameters are not used. The dr of assemblies 12 was conveniently measured by integrating the aminal methine proton, who's signal was well-resolved for almost all assemblies observed. The assemblies that did not feature two clear, well-resolved methine signals for the two diastereomers were not included in the plot. It should be noted that dr measurements were conducted using racemic 11a due to cost considerations. |
677dde8e6dde43c908acd364 | 0 | Deconstructive transformations in organic synthesis are valuable approaches to utilizing chiral pool materials that include functional groups that can be removed in a synthetic route. Performing deconstruction of Csp 3 -Csp 2 species can be particularly useful in terpene-based frameworks, as these chiral natural products often contain isopropenyl groups that are often not included in the final products. Hydrodealkenylation is a deconstructive transformation that enables the usage of chiral pool starting materials containing excess alkenyl functionality for the overall synthetic route. The ability to remove such groups without interfering with other functional groups already in place in the molecule is a challenge that has been explored briefly wherein ozonolysis is followed by Fe (II)-induced β-fragmentation to generate a number of derivatives (Fig. ) 1 . One caveat to this approach is the usage of ozone, as this electrophilic oxidant is known to react widely with a variety of functional groups. Generally, controlling the stoichiometry of O₃ for reaction with specific functional groups on substrates is challenging due to its gaseous nature. In contrast, under Mukaiyama-type conditions, optimizing reagent loading and catalyst structure allows for monoselective or regioselective functionalization of substrates bearing multiple olefins even in the presence of excess O2 2 . |
677dde8e6dde43c908acd364 | 1 | A key intermediate in ozonolytic cleavage of alkenes is the peroxyketal 3 . Access to peroxyketals can be achieved through ozone-based approaches, albeit this approach comes with some challenges. Ozonolysis on an industrial scale can be dangerous and impractical, although flow chemistry apparatuses can alleviate some of the challenges faced on scale. Additionally, unwanted reactivity of ozone with other oxidizable groups can limit reaction practicality in many systems. A synthesis of shearilicine by our group involves the hydrodealkenylation of an isopropenyl group, but such removal must be accomplished early in the route to avoid unwanted side reactivity with both the furan and indole moieties encountered later (Figure ) . The ability to remove isopropenyl fragments later would enable the broader use of dealkenylation as a general strategy in multistep synthesis. |
677dde8e6dde43c908acd364 | 2 | One approach to improving the selectivity of peroxide installation is the Mukaiyama peroxidation . Mukaiyama peroxidation and other Mukaiyama-type functionalizations can occur in the presence of indoles and electron-rich arenes, selectively functionalizing less electron-rich alkenes. Baran demonstrated the practical hydroamination of alkenes with nitroarenes, using Fe conditions on substrates with a variety of heterocycles present 6 . Renata chemoselectively functionalized an alkene with Mn(dpm)3 in the presence of lactam and methoxybenzene groups to generate a 5-deoxyterreulactone C precursor . Carreira performed an intramolecular cyclization/alkene hydration with cobalt as the catalyst, a method that tolerated the presence of a pyrrole . Boger utilized FeCl3 in the oxidation and functionalization of an alkene to form 20'vinblastine analogues, demonstrating selectivity of Fe catalyzed conditions to target alkenes over indoles . In route to the syntheses of Kopsia alkaloids, Magnus successfully oxidized an α,βunsaturated ester to the corresponding α-hydroxy ester using Mn(dpm)3 as a catalyst, in the presence of indoline heterocycles. Herzon performed intermolecular hydropyridylation with cobalt and pyridine salts, showing that pyridine-containing products would, in fact, be stable under Co catalyzed conditions . Within our group, Mukaiyama conditions were used to hydrate the alkene within a route to principinol D . Given the wide range of alkene substrates tolerant to Mukaiyama-type hydrations, there was an opportunity to develop a method of hydrodealkenylation that utilized such chemistry and would facilitate future multistep synthesis endeavors. We hypothesized that hydrodealkenylation could be accomplished utilizing Mukaiyama-type conditions for peroxide installation followed by Fe(II) promoted βfragmentation (Fig. ). |
677dde8e6dde43c908acd364 | 3 | Investigation began with the substrate 3a. Initially commonly employed hydroperoxidation conditions developed by Mukaiyama were used followed by treatment under a variety of reducing conditions (Fig. , Entry 1) . However, survey of a wide variety of conditions did not lead to the desired reduced compounds derived from β-scission. Ferrous sulfate, copper(II) tetrafluoroborate, and copper (II) sulfate, and others were investigated, with multiple unidentified products formed without the observation of the desired product. Various solvents (MeOH, DCE, DMF, i-PrOH, DCM, etc.) and combinations thereof were also examined. When silylhydroperoxidation was attempted followed by treatment under reducing conditions, the product could be observed. |
677dde8e6dde43c908acd364 | 4 | Multiple reductants were screened (Fig. ) after treatment with Co(thd)2 Et3SiH and O2 in DCE. The use of R-terpinene in substoichiometric amounts in the presence of MeOH did not lead to catalyst turnover (29% yield, Entry 2). Hantzsch ester and (R)-terpinene gave 57% and 65% yield respectively, comparable to the 65% observed with usage of PhSH (Entries 3-5). Without a hydrogen atom source present, only trace desired product was observed, with multiple unidentified adducts observable on TLC (entry 6). Many other metal reductants were explored that provided inferior results, including ferrocene, CuI, Mn(acac)3, and Ti(Cl)3. Telescoping the hydration and hydrodealkenylation reactions resulted in 71% yield of the desired product from the initial alkene, lending a one-pot procedure (Fig. ). |
677dde8e6dde43c908acd364 | 5 | One-pot Mukaiyama hydration paired with Fe(II) β-scission promotes the deconstructive manipulation of substrates containing isopropenyl groups. The ozone-free reductive dealkenylation protocol smoothly allows for the use of O2 instead of O3 for the peroxide formation step. Such a transformation may be highly enabling in multistep synthesis, such as shearilicine, wherein routes proceed via ozone sensitive substrates. |
677dde8e6dde43c908acd364 | 6 | To a solution of 3a (42 mg, 0.178 mmol, 1.0 eq) and triethylsilane (0.0567 mL, 0.356 mmol, 2.0 equiv.) in 1,2-dichloroethane (1.78 mL) in a test tube was added Co(thd)2 (3.78 mg, 0.0089 mmol, 5 mol %) at room temperature. The reaction mixture was stirred under an oxygen atmosphere (balloon) at room temperature for 1.5 h (The reaction mixture turned green after approximately 20 min). After the consumption of 3a on TLC (The corresponding peroxide appeared at a higher position on the TLC plate.), the solution was then sparged with nitrogen gas for 15 minutes to expel excess oxygen gas. MeOH (1.78 mL), γ-terpinene (0.114 mL, 0.712 mmol, 4 equiv.), and FeSO4•7H2O (69.3 mg, 0.249 mmol, 1.4 equiv.) were added sequentially to the reaction mixture at -30 °C. After stirring for 14 h at room temperature, the reaction was quenched by adding water. The reaction mixture was extracted with CH2Cl2 (three times), and the combined organic layers were dried over MgSO4, and then concentrated. After removing the solvent, column chromatography on silica gel with hexane/ethyl acetate (9/1 to 8/2) afforded the product 3c (25 mg, 71% yield) as a white powder. |
65bdfc829138d231615573d9 | 0 | Low temperature proton exchange membrane fuel cells (PEMFCs) powered by green hydrogen provide a means to sustainable energy conversion for sta�onary and transport applica�ons. Their widespread commercializa�on is par�ally limited by the cost of the pla�num (Pt)-based nanopar�cles supported on high surface area carbon (Pt/C) at the cathode, where oxygen reduc�on reac�on (ORR) occurs. Single iron (Fe), cobalt (Co), manganese (Mn) or �n (Sn) atoms (and their combina�ons) coordinated to nitrogen-doped carbon (M-N-C, where M is the metal) exhibit the most promising non-precious metal ac�vity for ORR. Of these, Fe-N-C has exhibited the greatest PEMFC performance. S�ll, ~60-100 µm Fe-N-C thick Fe-N-C cathodes are commonly used to compete with the PEMFC performance of ~5 µm Pt/C thick Pt-based cathodes, due mainly to Fe-N-Cs lower specific and volumetric ac�ve site density. With further improvements, Fe-N-C offer a poten�ally less expensive and less environmentally impac�ul alterna�ve to Pt/C, although highly ac�ve Fe-N-C typically suffer from lower durability. Researchers have improved the stability of Fe-N-C by improved synthesis pathways, producing atomically dispersed ac�ve sites, rather than encapsulated nanopar�cles, which induce instability. Most recently adding atomically thin protec�ve coa�ngs or reduc�ve pyrolysis condi�ons has led to Fe-N-C durability beyond 300 h in PEMFC under H 2 /Air. However, Fe-N-C durability is s�ll below commercial realiza�on for transport applica�ons (>5,000 h) owing to several degrada�on routes, which can be separated into two categories. Firstly, support modifica�on, such as oxida�on of the N-C matrix, and N-protona�on (especially for materials synthesized through pyrolysis under ammonia). Second is direct ac�ve metal atom modifica�on by agglomera�on/ aggrega�on, and demetalla�on/dissolu�on. The demetalla�on of the ac�ve site can also take place indirectly through chemical or electrochemical corrosion of the N-C matrix. Steps can be taken to deconvolute these degrada�on pathways and also minimize them, or even temporarily reverse them by reac�va�on. However, studies point towards the demetalla�on of FeN x ac�ve sites being the primary irreversible performance degrada�on mechanism in PEMFCs and the first step in the aggrega�on scenario. Induc�vely coupled plasma mass spectrometry (ICP-MS) is a highly sensi�ve technique which can provide �me-and poten�al-resolved Fe dissolu�on profiles from Fe-N-C catalysts. Monitoring Fe dissolu�on from ex situ ICP-MS, in tandem with other characteriza�on techniques, in rota�ng disc electrode (RDE) and PEMFC has revealed significant dissolu�on of Fe, although probing the mechanism requires operando measurements. In the first online flow cell ICP-MS study, Choi et al. suggested forma�on of insoluble ferric (Fe 3+ ) species, which dissolve under PEMFC opera�ng condi�ons (E cathode < 0.7 V RHE ) due to operando reduc�on to soluble ferrous (Fe 2+ ) ca�ons. This is in line with former ex situ ICP-MS findings of Zelenay and coworkers who suggested higher solubility of Fe species in acid solu�ons compared to Fe 3+ species. Previous online flow cell ICP-MS studies also provided cri�cal informa�on on the effects of pyrolysis atmosphere, bulk electrolyte pH, and catalyst modifica�on on the extent of Fe dissolu�on. Nonetheless, flow cell ICP-MS studies are limited to low current densi�es, and cannot reproduce all the prac�cal condi�ons occurring in an opera�ng PEMFC device (O 2 par�al pressure and current density, lower rela�ve humidity). In this respect, online gas diffusion electrode (GDE) ICP-MS is an adequate tool to simulate the environment of a PEMFC cathode more realis�cally, and gain PEMFCrelevant durability trends. For instance, Ehelebe et al. first demonstrated significantly lower dissolu�on of Pt/C catalysts in GDE configura�on compared to flow cell systems due to varying mass transport condi�ons of Pt species, 34 as previously proposed. Very recently, Choi and coworkers monitored in situ changes in ac�ve site density and operando Fe dissolu�on of a Fe-N-C under Ar and O 2 at different temperatures using GDE ICP-MS cell in acidic condi�ons. From site density monitoring, the reduced turnover frequency confirmed a reac�ve oxygen species catalyzed carbon corrosion scenario. However, despite using a GDE, Choi and coworkers current densi�es at 0.6 V chronoamperometric holds (<10 mA cm -2 geo ) were comparable to values achievable in flow cell (~1-2 mA cm -2 geo ), and not prac�cal PEMFCs. They observed from post-mortem transmission electron microscopy (TEM) and energy dispersive X-ray spectroscopy (EDXS) elemental mapping that Fe deposited as Fe x O y nanopar�cles a�er O 2 reduc�on in their Fe-N-C derived from microporous zeoli�c imidazolate framework-8 (ZIF-8), confirming earlier findings from Kumar et al. Evidence of Fe x O y nanopar�cle forma�on in PEMFC-relevant condi�ons has previously been ascribed to highly ac�ve but unstable high-spin FeN 4 C 12 moie�es, via Mössbauer spectroscopy. Temperature is also a cri�cal parameter for durability of Fe-N-C catalysts. Goellner et al. first evidenced that the rate of corrosion of a N-C matrix (150 square wave cycling between 0.9-1.4 V RHE , 3 s holds in RDE) increases 14-fold when temperatures increase from 20 to 80 o C. This resulted in 18-fold larger O 2 reduc�on ac�vity decay (at 0.8 V RHE ), which was assigned to N-C corrosion. Carbon corrosion can be avoided at 25 o C in RDE by keeping poten�al < 0.9 V RHE , although some carbon corrosion (<7 mA cm - 2 geo ) is reported in PEMFC at 80 o C. Kumar et al. reported Fe cluster forma�on under load cycling (Arsaturated 0.1 M H 2 SO 4 , 0.6-1.0 V RHE ) at 80 o C, but did not observe Fe clusters at 60 o C, providing strong evidence of the effect of temperature on the fate of Fe species. Finally, we note that Osmieri et al. reported greater performance loss under air-fed vs. N 2 -fed PEMFC cathode (3 s holds at 0.95 and 0.6 V vs anode , 80 o C), although with no nanopar�cle forma�on. There are thus conflic�ng results in literature, which is could be due to opera�on condi�ons, (temperature, gas atmosphere, Nafion content, current densi�es, poten�al etc.), storage condi�ons, electrode prepara�on and synthesized Fe-N-C proper�es. Moreover, most of Fe-N-C catalysts studied by operando ICP-MS have consisted of low ac�ve site u�liza�on Fe-N-C derived from ZIF-8. Our laboratory, and others, have highlighted that such catalysts display a predominantly or purely microporous structure. This limits the mass transport and electrochemical ac�ve site u�liza�on (number of electrochemically accessible FeN x sites to the total number of FeN x sites) to typically <10%. This prompted us to revisit Fe dissolu�on and the fate of Fe in FeN x ac�ve sites from our recently developed high FeN x u�lisa�on (>50%) Fe-N-C with high micro-and meso-porosity. This pore structure can facilitate mass transport of reactants for improved ac�vity, while also enabling transport of dissolved Fe ions for operando ICP-MS detec�on. |
65bdfc829138d231615573d9 | 1 | The impact on Fe-N-C ac�vity from changes in the (micro-) environment from RDE to GDE/PEMFC has been discussed. Local pH is one value which could vary between electrocatalysts and tes�ng condi�ons, and is recognized to influence Fe-N-C ac�vity. Meanwhile, the influence of pH on degrada�on is beginning to receive greater aten�on in modelling reac�on mechanisms and dissolu�on trends. Local pH (at the interface between the working electrode and the bulk of the electrolyte) and its effects has been inves�gated and discussed quite extensively in electrochemical CO 2 reduc�on; however, so far it has garnered limited experimental and theore�cal evidence for ORR. Meanwhile, kine�c modelling work by Zenyuk and Litster found during ORR increased pH along Pt mesopore channels, when devoid of Nafion and instead filled with water. It is worth considering that FeN x ac�ve sites are proposed to be located within micropores, which are expected to be filled with water. Even so, Banham and coworkers' experiments suggest that micropore flooding does not contribute significantly to PEMFC performance decay. Instead, kine�c models of Fe-N-C ac�vity decay under different poten�osta�c condi�ons in PEMFCs have been previously proposed, which has led to some debate. S�ll, to date these kine�c models of Fe-N-C have not factored in pH change and condi�ons in GDEs have not been considered. |
65bdfc829138d231615573d9 | 2 | In this work, we monitored Fe dissolu�on of a high electrochemical u�liza�on Fe-N-C catalyst using operando flow cell and GDE ICP-MS. We found that the fate of Fe-N-C catalysts is determined by combined Fe demetalla�on, reac�ve oxygen species ac�on (magnifying Fe demetalla�on) and local pH changes caused by ORR. We used a suite of complimentary preand post-mortem characteriza�on techniques (SEM, TEM, STEM, EDXS, EELS, Raman spectroscopy, XRD, XPS, XANES) to illustrate changes in structure and chemistry; based on our experimental insights, we built a microkine�c modelling to interpret our observa�ons. |
65bdfc829138d231615573d9 | 3 | The setup consisted of a homemade PEEK cell (Figure ) with a three-electrode configura�on using a glassy carbon rod as counter electrode (Sigradur grade G, HTW GmbH) and a leak-free Ag/AgCl/3.4 M Cl -(ET072, eDAQ) as reference electrode. The Ag/AgCl/3.4 M Cl -was calibrated versus reversible hydrogen electrode (RHE) via both a Hydroflex (Gaskatel) and a homemade Pt wire RHE. Fe in TAP 900@ Fe was used for online flow cell ICP-MS measurement to avoid interference from ArO + . The flow cell protocol and ICP-MS opera�on is detailed in the Supplementary Informa�on and Figure . |
65bdfc829138d231615573d9 | 4 | The GDEs were prepared by doctor-blade coa�ng an Fe-N-C ink onto a gas diffusion layer (GDL) including a microporous layer (Freudenberg, H23C8, 215.5±6.5 µm). During the doctor-blade coa�ng the temperature of the plate of automated film applicator (Zehntner, ZAA 2300) was at room temperature (23.5±0.5 o C). The composi�on of the ink was 12 wt.% solutes in a water/alcohol mixture, consis�ng of 68 wt.% isopropanol (Supelco, EMSURE, ACS ISO), 17.6 wt.% 1-Propanol, 13.6 wt.% water and <0.8 wt.% Ethanol, where the later three components are from the commercial Nafion solu�on (Fuel Cell Store, D2021, 21±1 wt.% Nafion, 34±2 wt.% water, 44±2 wt.% 1-Propanol, and < 2 wt.% Ethanol). The solute frac�on comprised 41.3 wt.% of TAP 900@Fe material and 58.7 wt.% of Nafion. Due to the high mesopore volume of TAP 900@Fe, a rela�vely high ionomer to Fe-N-C weight ra�o of 1.42:1 was used to ensure u�liza�on of the catalyst layer. A�er 30 min of s�rring and 1 h of sonica�on (100 W VWR Ultrasonic Cleaner USC 500 THD) at T < 30 °C, the ink was constantly s�rring un�l deposi�on. A�er the ink deposi�on onto the GDL, the samples were dried at room temperature (21±2 °C) under atmospheric pressure un�l tes�ng. The catalyst layer loading was 0.86±0.15 mg FeNC cm geo , as determined by weighing the GDE before and a�er Fe-N-C coa�ng. The catalyst layer thickness was 58±4 µm, as measured by a micrometer (Helios Preisser, 0912501). |
65bdfc829138d231615573d9 | 5 | Prior to electrochemical tes�ng, GDEs were immersed in ultrapure water for 1 hour. The electrolyte, reference and counter electrodes were 0.1 M HClO 4 (Suprapur, Sigma Aldrich), Ag/AgCl (inner and outer compartments filled with 3M KCl and 0.1 M HClO 4 , respec�vely, Metrohm) and Ti/Ir mixed oxide grid (METAKEM), respec�vely. Ag/AgCl/3M KCl was calibrated every day at the temperature of interest (E Ag/AgCl/Cl-= 0.316 ± 0.011 V RHE at 20 o C and E Ag/AgCl = 0.297 ± 0.013 V RHE at 75 °C). A gas humidifica�on system built with two gas washing botles (Duran) and a hea�ng plate (IKA TM RCT Basic Hot Plate S�rrer) was used to heat the purged gases to 75 °C. The GDE half-cell was heated to 74 ± 1 °C using an electrolyte recircula�on system via a hea�ng bath (AQUAline, LAUDA). In GDE, following the previously reported protocol, 64 100% post iR correc�on was applied for O 2 measurements, while for Ar measurements, 50% was applied in situ and 50% post Ar experiment. Details of GDE ICP-MS opera�on and protocol are detailed in the Supplementary Informa�on, Table and. The online Fe dissolu�on was measured with our previously reported GDE ICP-MS setup, shown in Figure . |
65bdfc829138d231615573d9 | 6 | A one-dimensional model was developed to describe pH distribu�on in the catalyst layer. This model encompasses a system of par�al differen�al equa�ons (Supplementary Informa�on) that account for the transport of Fe and protons in the electrolyte and the 60 µm thick catalyst layer, as well as the proton consump�on by the ORR and the dissolu�on/precipita�on of Fe ca�ons in the catalyst layer. The modelling is based on the following assump�ons: |
65bdfc829138d231615573d9 | 7 | Due to the pronounced difference in complexa�on constants, only Fe 3+ ca�ons are expected to precipitate. If Fe 2+ ca�ons are dissolved in water, they will anyway thermodynamically be oxidized into Fe 3+ ca�ons by O 2 . (iv) Based on the GDE ICP-MS data at 20 °C that will be discussed later, the rate of produc�on of dissolved Fe ions is assumed to be approximately two �mes faster in O 2 than in Ar GDE experiments. (v) A homogeneous poten�al distribu�on is assumed in the catalyst layer. |
65bdfc829138d231615573d9 | 8 | Comparing TAP 900@Fe and TAP 900@ 57 Fe RDE ORR Ac�vity Thorough ex situ characteriza�on of TAP-derived materials was carried out in our previous work. Although some comparisons between TAP 900@ 57 Fe and TAP 900@Fe were missing. Considering O 2 reduc�on, reduced ac�vity has previously been reported for Fe enriched Fe-N-C samples compared to Fe-N-C prepared in the same manner but with natural abundance Fe precursor. The RDE O 2 reduc�on mass ac�vity for TAP 900@ 57 Fe and TAP 900@Fe can be found in Figure . The kine�c region and mass ac�vity at 0.8 V RHE, iR-free in O 2 -saturated RDE is lower in TAP 900@ Fe compared to previously reported TAP 900@Fe, with 3.77±0.54 and 5.01±0.79 A g FeNC -1 , respec�vely (Figure ). The lower ac�vity with Fe enrichment follows the previous report. |
65bdfc829138d231615573d9 | 9 | Moving to operando flow cell ICP-MS measurements in 0.1 M HClO 4 , TAP 900@ 57 Fe was used to avoid polyatomic interference from ArO + and maximize spectrometric signal. The setup and experimental protocol are depicted in Figure and Figure , respec�vely. First, ICP-MS calibra�on, electrochemical impedance spectroscopy and open circuit poten�al (OCP) measurements were conducted to ensure correct installa�on and opera�on. Next, 50 fast (50 mV s -1 ) cyclic voltammograms (CV) between 0.925-0.200 V RHE were measured in Ar-saturated electrolyte to allow the catalyst to reach a stable electrochemical and dissolu�on measurement (Figure , 0.2 mg FeNC cm -2 geo ). Mg was also monitored during the ini�al 50 cycles due to its use as a templa�ng agent during synthesis, with 0.06 wt.% detected from ex situ ICP-MS in our previous work. Mg dissolu�on did not vary with poten�al (Figure ) and so is not considered further. Meanwhile, the rate of Fe dissolu�on followed an exponen�al decay. |
65bdfc829138d231615573d9 | 10 | Considering the effect of increased Fe-N-C loading, the amount of Fe dissolu�on follows a linear trend over the ini�al 50 CVs (Figure ). The percentage of total Fe detected increases from 7.5±2.9% to 15.2±3.3% as catalyst loading increases from 0.05 to 0.40 mg Fe-N-C cm geo , with 11.3±5.6% at 0.20 mg Fe-N-C cm geo (Figure ). This finding appears counterintui�ve as one would expect either an equivalent percentage of Fe detected rela�ve to the loading, or even a reduced percentage of detected Fe, due to reduced ac�ve site u�liza�on with increasing thickness of the catalyst layer. It is also worth no�ng that there is a constant 130 ng Fe g FeNC -1 s -1 57 Fe concentra�on observed when held at 0.9 V RHE (Figure ), which was also the OCP of the TAP 900@ 57 Fe catalyst. |
65bdfc829138d231615573d9 | 11 | A�er the ini�al 50 CVs at 50 mV s -1 , six CVs were conducted at 10 mV s -1 under Ar and then O 2 -satura�on. With increasing TAP 900@ 57 Fe loading under O 2 -satura�on, the limi�ng current density (below 0.65 V RHE ) only incrementally increases. This slight increase can be explained by the increasing thickness of the catalyst layer with loading, which penetrates deeper into the flowing O 2 -saturated electrolyte. Meanwhile, between 0.65-0.80 V RHE there is an increasing O 2 reduc�on peak in the cathodic direc�on (Figure ). This is caused by a build-up of O 2 concentra�on locally in the catalyst layer while scanning the poten�al region of 0.800-0.925 V RHE , where very litle ORR is observed. |
65bdfc829138d231615573d9 | 12 | Under Ar-saturated condi�ons the current density increases propor�onal to the catalyst loading; we note there is an increasing peak on the cathodic scan (Figure ). We atribute this cathodic peak to the reduc�on of trace O 2 , arising from air ingress at the junc�on of the Kalrez O-ring and cell (or cavita�on from the peristal�c pump). S�ll, the amount of O 2 appears negligible. Normalizing the Fe detected to charge passed and catalyst loading shows the amount of Fe detected is constant under O 2 but increases with reduced catalyst loading under Ar (Figure ). Meanwhile, the amount of Fe detected is equivalent under either gas satura�on, with 1.3-2.0% of total Fe detected, and linear dependence with Fe-N-C loading (Figure ). Focusing on the dissolu�on at 0.2 mg FeNC cm geo , similar profiles are observed under Ar and O 2 -satura�on (Figure ). |
65bdfc829138d231615573d9 | 13 | To evaluate differences in Fe detec�on and profiles over a longer period, 1 h AST or chronoamperometry (CA) were recorded in 0.1 M HClO 4 (Figure ) or 0.05 M H 2 SO 4 (Figure ). Greater Fe loading-normalized concentra�on is observed over the course of the AST under O 2 than Ar. Fe concentra�on follows a slow decline under O 2 and rapid plateau above baseline under Ar (Figure ). 2.7±0.1% of total Fe is detected during O 2 AST (Figure ), with a charge normalized Fe dissolu�on of 503±3 ng Fe mg FeNC -1 C -1 (Figure ). Meanwhile, half Fe concentra�on is observed under Ar AST (Figure , Figure ); however, normalizing to the total charge passed shows approximately double, with 1022 ng Fe mg FeNC -1 C -1 (Figure ). Preand post-mortem bright-field TEM of these samples shows no forma�on of detectable nanopar�cles under Ar or O 2 (Figure ), indica�ng all Fe demetalla�on leads to dissolu�on at 25 o C, in agreement with former findings of Kumar et al. CA under O 2 at 0.2 V RHE shows a large ini�al spike in Fe concentra�on, which then decays over �me, while CA at 0.6 V RHE shows a smaller spike and lower overall dissolu�on (Figure ). The ini�al spike in 57 Fe concentra�on may be related to double layer charging and rapid change in poten�al. A�er 30 mins, the current density and Fe dissolu�on are equivalent at 0.2 and 0.6 V RHE CA. CA at 0.2 V RHE ends with 4.6±0.4% of total Fe and 686±166 ng Fe mg FeNC -1 C -1 . This is approximately double the values at 0.6 V RHE , with 2.2±0.1% Fe and 358±61 ng Fe mg FeNC -1 C -1 (Figure ). This correlates with the observa�ons from Figure , where greatest Fe dissolu�on occurs around 0.20 V RHE . |
65bdfc829138d231615573d9 | 14 | At the same pH in 0.05 M H 2 SO 4 instead of 0.1 M HClO 4 , O 2 AST shows a similar dissolu�on profile, with lower Fe detec�on but higher O 2 current densi�es (Figure ). This difference in current is unexpected as O 2 solubility is comparable at these acid concentra�ons. Meanwhile the total Fe loss is 2.3±0.1% in 0.05 M H 2 SO 4 and slightly higher in 0.1 M HClO 4 with 2.7±0.1% (Figure ). However, the charge normalized Fe dissolu�on is less than half in 0.05 M H 2 SO 4 , at 231±63 ng Fe mg FeNC -1 C -1 (Figure ). |
65bdfc829138d231615573d9 | 15 | While flow cell measurements proved insigh�ul, the degrada�on rate in real PEMFCs may be different due to higher ORR rates and therefore ORR charge passed. To reach higher current densi�es and condi�ons comparable to PEMFCs, TAP 900@Fe was tested in a GDE half-cell coupled to online ICP-MS in 0. geo (Figure ). This is assigned to improved we�ng of TAP 900@Fe during the 20 o C Ar AST. Meanwhile, 20 o C O 2 AST led to no�ceable performance degrada�on a�er only 200 cycles, with poten�al shi� at -50 mA cm geo of -50±30 mV (from 0.61±0.03 to 0.56±0.00 V RHE, iR-free ) compared to pris�ne 20 o C TAP 900@Fe (Figure ). At -50 mA cm geo , 75 o C GDE pris�ne TAP 900@Fe shows an ), with slight reduc�ons in C=O and C-O peaks for AST samples (Figure ). Meanwhile, a�er 75 o C O 2 protocol, a clear overall O1s increase is found, equivalent to 12.2 at.% O1s (Figure ). There is less discernible change in the C1s spectra, aside from reduc�on in C-N and C-C and increase in CF2 in all AST samples compared to the pris�ne TAP 900@Fe GDE (Figure ). Raman spectra (Figure Within the pris�ne TAP 900@Fe GDE no visible nanopar�cles are detected using HAADF-STEM and STEM-EDXS spectrum imaging (Figure , Figure ). A�er 20 o C Ar protocol, one large Fe x O y nanopar�cle is detected in the spectrum image, while, at higher magnifica�on, small clusters are observed (Figure , Figure ). Numerous Fe nanopar�cles are observed following 20 o C O 2 protocol in GDE, which are assigned to Fe x O y based on overlaying the Fe and O EDXS mapping (Figure , Figure ). HAADF-STEM combined with EDXS and EELS reveals clusters containing Ca and Fe in fresh and post Ar and O 2 AST GDE (Figure ). The presence of Ca remains unexplained, as we consistently used MQ water for all our electrochemical experiments and rinsing steps. No trace of Ca was also detected in the na�ve catalyst. We therefore atribute it to contamina�on by tap water. The peak at 695 eV is from Fe-K. STEM-EELS analysis in regions without Fe par�cles cannot resolve any Fe peak (Figure ), likely owing to the concentra�on of FeN x sites being below the limit of detec�on. |
65bdfc829138d231615573d9 | 16 | Post 75 o C O 2 protocol no large Fe x O y par�cles are seen from EDXS and limited Fe clusters from HAADF-STEM (Figure , Figure ). No significant change from the pris�ne TAP 900@Fe structure is observed a�er 20 o C O 2 and Ar protocols, (Figure ); however, a�er 75 o C O 2 protocol a denser par�cle structure is observed (Figure ). |
65bdfc829138d231615573d9 | 17 | XRD on post-mortem GDE AST samples was conducted to try and deduce the type of Fe x O y , however either the lack of crystallinity, small par�cle size and/or low concentra�on meant no sharp peaks rela�ng to Fe par�cles could be iden�fied (Figure ). The peak at 18.0 o is assigned to polytetrafluoroethylene, which arises from the Nafion backbone. It is worth men�oning that pris�ne TAP 900@Fe does not show a graphite peak at ~25.6 o (002), sugges�ng its amorphous or graphenelike structure, with an average of single atomic layers found from previous Raman analysis. Normalized absorp�on and first deriva�ve XANES of fresh TAP 900@Fe powder and GDE ink, plus post Ar and O 2 25 o C protocols, are compared to references of Fe foil, FeO and Fe 2 O 3 in Figure . A posi�ve shi� of center of mass of the pre-edge in TAP 900@Fe ink and a�er Ar and O 2 protocols signifies an increase of oxida�on state, while their decrease in intensity is related to a change in local coordina�on of Fe. TAP 900@Fe GDE ink displays a near iden�cal spectra to post 25 o C O 2 . This suggests changes in Fe coordina�on and oxida�on state between TAP 900@Fe powder and its ink, similar to a recent report by Saveleva et al. for other Fe-N-Cs. Post Ar protocol shows a lower rising edge posi�on indica�ng a lower average Fe oxida�on state, or change in bond length and/or coordina�on change. |
65bdfc829138d231615573d9 | 18 | To elucidate the Fe dissolu�on mechanisms in a prac�cal device, online GDE ICP-MS was measured before, during and a�er the AST (Figure ) for each of the condi�ons. It is observed that the baseline Fe concentra�on is high even a�er the preliminary 50 CVs in Ar (50 mV s -1 ). |
65bdfc829138d231615573d9 | 19 | For ini�al 20 o C Ar (Figure ), Fe concentra�on above baseline occurs at 0.83 V RHE, iR-free and reaches a maximum concentra�on between 0.64-0.48 V RHE, iR-free . Returning to 0.83 V RHE, iR-free , Fe concentra�on returns to baseline levels. Ini�al 20 o C O 2 current step holds (Figure ) show a lower baseline Fe concentra�on than 20 o C Ar. A fall in Fe concentra�on below baseline levels is observed when increasing current density from -1 to -15 mA cm -2 geo , corresponding to 0.85±0.02 to 0.80±0.01 V RHE, iR- free , respec�vely. When returning anodically to hold at -1 mA cm geo , Fe concentra�on increases and only begins falling back to baseline once returning to hold at -0.05 mA cm geo . Ini�al 75 o C O 2 current hold measurements show a higher baseline Fe concentra�on, with increased Fe concentra�on during holds at -1 to -15 mA cm geo . Fe concentra�on then returns to approximate baseline values during holds at -50 and -100 mA cm geo , corresponding to 0.66±0.04 and 0.62±0.04 V RHE, iR-free , respec�vely. Returning anodically to holds at -15 and -0.6 mA cm geo results in increased Fe concentra�on. |
65bdfc829138d231615573d9 | 20 | Our experiments in flow cell ICP-MS first confirm that the dissolu�on of Fe atoms is indeed the predominant degrada�on mechanism in this type of catalyst. In CV, two well-defined Fe dissolu�on peaks can be observed, with onset of 0.73 and 0.33 V RHE on the cathodic scan (Figure ). The two Fe concentra�on peaks could represent two different Fe species dissolving at different poten�als, or different dissolu�on process with different formal poten�als. Only one Fe concentra�on peak was resolved by Santori et al., with an onset of Fe concentra�on at ca. 0.75 V RHE for their Ar-pyrolysed Fe-N-C in O 2 -saturated 0.1 M H 2 SO 4 at 2 mV s -1 (data reproduced in Figure ). Meanwhile Choi et al. observed the onset of increased Fe concentra�on at 0.77 V RHE , with two dis�nguishable Fe dissolu�on peaks, as observed here. The poten�al at which peak Fe concentra�on occurs is not discussed as this depends on mass transport, which changes with the electrochemical cell design and opera�ng condi�ons. |
65bdfc829138d231615573d9 | 21 | We note that the onset of increased Fe concentra�on at ca. 0.73 V RHE on the cathodic scan (Figure ) coincides with the onset of the quinone-hydroquinone redox on the cathodic scan post 8,000 O 2 AST at 80 o C (Figure ), and the second peak onset of increased Fe concentra�on at 0.33 V RHE on the cathodic scan coincides with the onset of the second reversible redox on the cathodic scan. Our observa�ons suggest that the stability of the Fe centre may be intrinsically linked to the chemistry of the surrounding ligands; this no�on is analogous to rela�onships observed by others between the cataly�c ac�vity and the chemistry of the surrounding ligands. We also note an ini�al exponen�al decay in Fe concentra�on (Figure ), which was also observed by Choi et al. for their Fe-N-C catalyst. In our case maximum Fe concentra�on is observed instantaneously upon poten�al cycling in Ar, whereas in the report of Choi et al. maximum Fe concentra�on is reached a�er 2-3 CVs. This could be due to the vastly different catalyst structures between our highly micro-and mesoporous TAP 900@ 57 Fe with high ac�ve site u�liza�on, and the bulky par�cle and predominantly microporous ZIF-8 derived Fe-N-C of Choi et al. Alterna�vely, it could arise from mass transport effects from slow residence �me in Choi et al.'s flow cell design. The structure of unmodified microporous ZIF-8 derived materials will have impeded mass transport, low ac�ve site u�liza�on and therefore delayed detec�on of Fe dissolu�on. Differences in experimental setup and residence �me calibra�on in this work and that of Choi et al. could also contribute to the observed �me difference in Fe concentra�on detec�on. |
65bdfc829138d231615573d9 | 22 | Choi et al. detected ~3% of total Fe over their ini�al 20 CVs in Ar-satura�on (100 mV s -1 , 0.8 mg Fe-N-C cm -2 ) for their purely microporous ZIF-derived Fe-N-C. While a�er 50 CVs (50 mV s -1 , 0.4 mg Fe-N-C cm -2 ), TAP 900@ 57 Fe shows 15.2 ± 3.3% Fe detected. This again points to the different porosity and structure in TAP-and ZIF-derived materials, leading to different accessibility of Fe sites. Although, it should be noted, according to our previous ex situ TAP900@ 57 Fe Mössbauer assignments, ca. 11% of the Fe existed as inac�ve FeCl 2 •4H 2 O. This species may represent some or all of the ini�ally dissolved Fe species. |
65bdfc829138d231615573d9 | 23 | Results from Figure (and Figure ) suggest that the Fe concentra�on is independent of O 2 reduc�on under cyclic voltammetry (0.9-0.2 V RHE at 10 mV s -1 ) in flow cell. This is contrary to what is observed in Figure , where detected Fe concentra�on is greater under O 2 than Ar under AST (step from 0.9 to 0.6 V RHE with 3 s poten�al holds) flow cell condi�ons. These difference observa�ons of Fe concentra�on may be due to either the different poten�als scanned (AST: 0.9-0.6 V RHE versus CV: 0.925-0.2 versus), the poten�al scanning protocol (AST: 3 s square wave voltammetry holds versus CV: 10 mV s -1 ), or 6 CVs not providing enough cycles to dis�nguish changes in Fe concentra�on. |
65bdfc829138d231615573d9 | 24 | Considering Figure and Figure , Zelenay and coworkers also observed from ex situ ICP-MS that HClO 4 dissolved more Fe from their polyaniline-derived Fe-N-C than H 2 SO 4 , which they atributed to differences in solubility of Fe perchlorates and sulfates. We suggest this observa�on could also be atributed to the stronger SO 4 2-binding on the Fe site, whereas ClO 4 -has been proposed to mimic non-specifically adsorbing proper�es of perfluoro sulfonic acid ionomers. If true, this would imply AST measurements in H 2 SO 4 in RDE and GDE would lead to slower FeNC degrada�on than in HClO 4 (at the same pH), when Fe dissolu�on is the main degrada�on mechanism. |
65bdfc829138d231615573d9 | 25 | The different current density under 0.1 M HClO 4 and 0.05 M H 2 SO 4 (Figure and) may be related to kine�c effects of the proton donor. Addi�onally, at 0.8 V RHE Fe-N-Cs have recently been reported to possess 1.3-2.9 higher mass ac�vity in H 2 SO 4 than HClO 4 . 71 |
65bdfc829138d231615573d9 | 26 | Under ini�al Ar in GDE ICP-MS (Figure ), the most significant increase in Fe concentra�on occurs when the poten�al drops from 0.83 to 0.74 V RHE . This can be explained by the Fe 3+ /Fe 2+ redox transi�on at 0.76 V RHE (Figure ). It is worth no�ng that with a Fe-N-C, Fe atoms possess different formal redox and dissolu�on poten�als depending on their coordina�ng ligands and extended local environment (number and size of graphene sheets, oxygen func�onal groups ). This broad Fe 3+ /Fe 2+ redox range is also ini�ally observed in Figure . Moreover, a�er the increases of Fe concentra�on during cathodic poten�al shi�s, gradual declines in the Fe concentra�on are frequently observed. This is related to the fact that the loca�on of the Fe within the Fe-N-C structure (outer catalyst layer surface or deeper within) affects the transfer func�on and hence residence �me. |
65bdfc829138d231615573d9 | 27 | It has been previously suggested that O-containing groups on the carbon surface reduce the turnover frequency of Fe-N-Cs by weakening O 2 -binding on FeN x sites. The limited change in XPS O1s spectra between pris�ne and 20 o C O 2 GDE protocol (Figure ) suggests performance degrada�on from 20 o C O 2 protocol (Figure ) is mainly atributed to ac�ve site demetalla�on. Meanwhile, the increase in O1s a�er 75 o C O 2 protocol (Figure ) causes reduc�ons in TOF and FeN x sites' stability and the increased observa�on degrada�on. Reduc�on in TOF occurs due to reac�ve oxygen species catalyzing mild carbon corrosion. The rapid decay in O 2 reduc�on performance (e.g. -50±30 mV at 50 mA cm geo a�er 200 cycle AST in 20 o C O 2 ) and high Fe dissolu�on can be atributed to the high percentage of unstable high spin Fe 3+ N x present (assuming the same type of sites are present between TAP 900@ 57 Fe and TAP 900@Fe). Addi�onally, according to density func�onal theory (DFT) calcula�ons for Fe-N-C, the number and size of graphene sheets affects the Fe dissolu�on poten�al. Previous Raman analysis of TAP 900 determined an atomically thin carbon structure, which therefore leads itself to possess less stable FeN x sites. |
65bdfc829138d231615573d9 | 28 | The online GDE ICP-MS concentra�on profile under 20 o C O 2 (Figure and) suggests Fe dissolu�on and subsequent detec�on by ICP-MS at low current density (-0.05 to -1.00 mA cm -2 geo ). Meanwhile, at higher current density (-15, -50 and -100 mA cm -2 geo ), a process of Fe dissolu�on and redeposi�on locally into Fe x O y in the catalyst layer is proposed. This is supported by the increased observa�on of Fe x O y a�er O 2 GDE protocol from HAADF-STEM and EDXS (Figure and Figure ). The reason for Fe x O y forma�on is hypothesized to arise based on the Fe Pourbaix diagram, where an increase in the local pH would form Fe 2 O 3 . This pH increase in the catalyst layer could occur due to the rapid consump�on of H + during increased O 2 reduc�on currents (4H + + O 2 + 4e -2H 2 O). It is then expected that some Fe x O y redissolves when returning anodically to low O 2 reduc�on current density (-1 mA cm - 2 geo ), due to a return to acidic pH. This redissolu�on is evidenced by the detected increase in Fe concentra�on at -1 mA cm -2 geo on the anodic step for 20 o C O 2 in GDE ICP-MS. The observa�on of Fe x O y corroborates previous findings from post-mortem O 2 AST protocols. Moreover, the increased Fe concentra�on detected when stepping the poten�al down in the cathodic direc�on a�er post AST (Figure ) for O 2 GDE at 20 o C and 75 o C supports the hypothesis that Fe x O y builds up in the catalyst layer at current densi�es of -50 mA cm geo during the AST and is only released at lower current density holds (-1 mA cm geo at 20 o C and 75 o C). |
65bdfc829138d231615573d9 | 29 | Mass transport (O 2 solubility and H + ) and the thermodynamics and kine�cs of ORR and Fe dissolu�on (at a constant poten�al on the RHE scale) will all change with temperature. This makes it challenging to deconvolute their contribu�ons to changes in performance; however, kine�c modelling based on experimental data can help explain phenomena, such as local pH changes. |
65bdfc829138d231615573d9 | 30 | We developed a microkine�c model of the system (Figure and Suppor�ng Informa�on) to replicate the observa�ons from GDE ICP-MS prior to AST at 20 o C in 0.1 M HClO 4 and evidence our hypothesis on the pivotal role of local pH. The model assumed the ini�al proton concentra�on and poten�al in the catalyst layer is homogeneous. We focus on the Fe concentra�on observed in GDE ICP-MS at 0.75 V RHE, iR-free and 20 °C, corresponding to a current density of -15 and 0 mA cm -2 geo under O 2 and Ar supply, respec�vely. The void volume (ε) in the catalyst layer was adjusted to semi-quan�ta�vely simulate the �me evolu�on of the Fe concentra�on signal monitored by GDE ICP-MS in Ar-saturated electrolyte (Figure and). The value of the proton consump�on rate constant (k r ) and ε were then varied to replicate the Fe concentra�on signal measured in O 2 -saturated electrolyte (Figure ). Good agreement between experiment and simula�on are reached for the range of values considered (0.2≤ ε ≤0.4 and 100≤ k r ≤400 s -1 ). Addi�onally, values for the tortuosity factor, τ (=1/√ε) were within previously reported ranges (1.8 ≤ τ ≤2.2). Figure displays the corresponding simulated pH profile in the catalyst layer. The simula�ons predict a significantly lower concentra�on of detected Fe ca�ons during O 2 reduc�on (Figure ). This phenomenon is atributed to the precipita�on of Fe 3+ ca�ons under the local condi�ons in the catalyst layer, with the Fe concentra�on resul�ng from the balance between Fe precipita�on and redissolu�on. Indeed, simula�ons indicate that at -15 mA cm -2 geo , the local pH at the interface between the Fe-N-C layer and the electrolyte solu�on is approximately 1.5 (Figure ). There is then a substan�al and rapid increase in pH moving into the bulk catalyst layer (far from the liquid electrolyte), reaching pH values ca. 1. The fate of Fe (and other metal species) can vary through the catalyst layer and should be considered when conduc�ng operando and post-mortem studies. |
65bdfc829138d231615573d9 | 31 | 2. Precious metal-free layers in PEMFCs, which typically employ 60-100 µm M-N-C thick cathodes, may not u�lize the majority of the catalyst layer during O 2 reduc�on due to proton consump�on. Focus, therefore, should be made on decreasing the electrode thickness by further increasing the electrochemically accessible volumetric ac�ve site density of precious metal-free catalysts. |
63ce9d531fb2a87c4addbfaa | 0 | Flow chemistry enables reactions to be performed as a continuous process with precise control of reaction conditions such as temperature, pressure, and stoichiometry. Flow reactions can be performed in two regimes -continuous or slug flow. In continuous flow the entire fluidic system is filled with reagents, intermediate or formed product. Conversely in slug flow much of the fluidic system is filled with carrier solvent, small discrete reaction slugs then flow through the system (Figure ). Although slug flow reactions are being performed on a much smaller reaction scale, on a microfluidic scale the same reaction is performed. This means at identical positions within the flow reactor an identical profile is observed in both the slug and continuous flow regime. This linear scalability makes slug flow chemistry ideal for library synthesis in flow. |
63ce9d531fb2a87c4addbfaa | 1 | Two critical considerations when performing slug flow chemistry is the timing and profile of the slug. Knowing to a high degree of accuracy where a reaction slug is within a fluidic system that is likely 10 or more times the volume of the reaction slug is crucial to enable effective collection of the product solution. This becomes even more apparent when two reactant slugs need to meet at a t-mixer. For a complete reaction, both slugs need to begin and end with perfect synchronicity (Figure ) -where the two slugs do not overlap one reagent does not see the other, meaning no reaction can occur (Figure ). |
63ce9d531fb2a87c4addbfaa | 2 | Prediction of slug timing within a simple tubular flow reactor can be achieved when both the flow rate and system volume are known to a high degree of accuracy, so long as the flow rate is well controlled. In more complex fluidic systems however -in particular when a solid phase reactor is present -prediction of slug timing can become more difficult to predict due to the chromatographic effect of the stationary phase reaction surface. Figure |
63ce9d531fb2a87c4addbfaa | 3 | If reaction slugs are highly coloured, and see-through PTFE/PFA or similar material tubing is used then slug timing can be corroborated by eye. If either of these factors is not the case however, then inline detection is required to 'see' the reaction slug. Several inline detection technologies are commercially available including inline flow NMR, MS, IR or UV. All of these commercial examples however are significantly priced and have a sizeable footprint to accommodate within a fume cupboard that likely already contains a large amount of equipment associated with the flow reactor itself. As a result, multiple groups have reported their efforts to overcome these issues, creating low-cost inline detectors relying on the absorbance of visible spectrum Whilst elegant and likely to generate high quality and sensitive data thanks to the reference measurement being taken, the resultant complex optical setup using a beam-splitter cube and optical fibres of the detector published by Glotz & Kappe makes replication of the literature challenging without sufficient engineering expertise. |
63ce9d531fb2a87c4addbfaa | 4 | The later publication from Höving et al. was the basis of a much simpler design. In the Höving et al. publication, the flow reactor tube was sandwiched between a coloured LED (Light Emitting Diode) and LDR (Light Dependent Resistor) and held together in a specially designed 3D printed housing. The simplicity of this design made it a much more tangible prospect for us to be able to reproduce. The reliance of the design however, on a single wavelength LED, and typically low sensitivity LDR, made us question if further enhancements could be made to increase the applicability. |
63ce9d531fb2a87c4addbfaa | 5 | With the rise of single board computers such as the Raspberry Pi and Arduino Uno a market has emerged for a wide range of sensors that can easily be incorporated into these systems. For this application, the Adafruit AS7341 sensor looked ideal. The sensor measures light across 10 wavelength 'channels' (IR, 8 visible spectrum wavelengths, and a 'clear' channel that covers that entire spectrum) and can be interfaced with easily using I 2 C (Inter-Integrated Circuit). If a white LED was used to irradiate the reactor pipe, then the absorbance of the fluid within the pipe could be determined at multiple wavelengths simultaneously -providing richer data than the previously published inline detectors at a similar low cost. |
63ce9d531fb2a87c4addbfaa | 6 | Using the "Version B" flow cell design from the Höving et al. publication as a starting point, an initial prototype detector housing was designed in Tinkercad. Although only 2.6 x 1.8 cm in size, the Adafruit AS7341 sensor board is significantly larger than an LDR as used in the previous publication. As a result, modifications were required to encase the larger Adafruit AS7341 sensor board to create a lightproof box, whilst exposing the sensor pins to allow circuit prototyping with the Raspberry Pi and a breadboard (Figure ). |
63ce9d531fb2a87c4addbfaa | 7 | To complete the prototype device a white LED was also controlled by the Raspberry Pi, using PWM (Pulse Width Modulation) to optionally control brightness of the LED, allowing the effect of varying irradiation brightness to be tested. Initial experiments using this prototype device confirmed our hopes. Using the example code provided by Adafruit and pumping test solutions through the flow cell using a hand operated syringe; THF, IPA (isopropyl alcohol) and ethyl acetate showed differences in absorbance, despite all being colourless liquids to the human eye. |
63ce9d531fb2a87c4addbfaa | 8 | To transform the prototype into a useable device that could be housed and operated in a laboratory fume cupboard two major improvements needed to be made: the fragile breadboard prototype circuitry needed to be rigidified and then protected from potential solvent splashes etc, and a better representation of the sensor data also needed to be created rather than using the Adafruit example code. |
63ce9d531fb2a87c4addbfaa | 9 | To consolidate and rigidify the circuitry bespoke printed circuit boards were designed using circuit design software tool Fritzing and then fabricated at a low volume PCB manufacturer, Aisler. A design strategy commonly employed in the single board computer arena was used -a Pi HAT (Hardware Attached on Top) circuit board was designed to contain all the control electronics in an easily attachable/detachable board. A second smaller circuit board was designed for the detector and LED. |
63ce9d531fb2a87c4addbfaa | 10 | On both boards was also included an RJ45 (ethernet) port. An ethernet cable was chosen to consolidate the 6 wires required to power the LED and communicate with the sensor into a single cable (leaving two of the eight wires unused). The choice was made to use a ubiquitous, easily interchangeable cable so that any cable of suitable length could be selected as required dependent on location. This design then allows the detector and Raspberry Pi to be located any distance apart by simply choosing a longer or shorter ethernet cable (Figure ). Further testing of the device was carried out to determine the best operation protocol. It was found slugs were much easier to visualise graphically if the sensor channel data is baselined to the absorbance of the system solvent (see Supplementary Information, Figure ). This then results in a clear positive or negative peak in the trace when a slug is passing the detector. LED brightness was also tested, and full power was found to give the best signal to noise ratio (the LED is not too bright to cause saturation of the detector). |
63ce9d531fb2a87c4addbfaa | 11 | Finally, we needed to provide a secure and protective housing for the new detector circuit board and separately the Raspberry Pi computer. The initial prototype 3D printed sensor housing was modified further to accommodate the RJ45 connector now included. The sensor device is an approximately 4.5 x 4.5 x 4.5 cm cube -still a very small footprint and thus very portable, allowing positioning wherever required in a fluidic system (Figure ). For the Raspberry Pi housing, a suitable case available on the Thingiverse was chosen. Minor modifications were made to accommodate the RJ45 connector we introduced, which stands taller than most Pi HAT devices. |
63ce9d531fb2a87c4addbfaa | 12 | With construction of the final design complete we carried out a more thorough validation test using our Syrris Asia flow chemistry system. Running THF as the system solvent we injected sequential slugs of dilute iodine in THF (as a coloured, positive control) followed by IPA and then ethyl acetate. All slugs could be clearly visualised as distinctly different to the THF system solvent -confirming the device works as anticipated (Figure ). |
63ce9d531fb2a87c4addbfaa | 13 | In conclusion, in this publication we report a significant enhancement on the previously published low-cost inline detector from Höving et al. by changing the detector to one that can measure multiple wavelengths simultaneously and changing the irradiation source to a white LED. As a result, we were able to obtain real-time inline spectral data from a flow reactor operating in a slug flow regime, indicating when slugs pass the detector even when undetectable to the human eye. This device enables slug timing within a flow reactor to be determined in a quick and facile manner. Furthermore, the device could be used to monitor residence time distribution within a system based on shape of the slug peak, along with perhaps estimation of reaction conversion if the optical absorbance profile of starting material and product were known. Details to construct the device are provided in the supplementary information document, including links to purchase the printed circuit boards directly from the manufacturer we used. Source code, 3D print .stl files and circuit board Gerber files (if you wish to organise manufacture yourself) are all provided on our GitHub repository: |
610936ae8804439fdde55e7b | 0 | Metal-organic framework (MOF) chemistry has flourished through the creation of a vast chemical space where more than 100,000 MOFs have been discovered. The number is increasing rapidly with a wide and continuously expanding variety of structural types, building units, linkage chemistry, and functional groups. In fact, the chemical space of possible MOF structures is so huge that it is impossible to fully explore it experimentally. , Simulation and machine learning (ML) have evolved as important tools for guiding researchers to computationally identify regions of interest. However, in order to synthesize the novel MOF structures, the researchers still have to rely on their experience, employing a trial-and-error approach (Fig. ). This is a very challenging process that is highly time-consuming, labor-intensive and requires a lot of resources. Therefore, the search for an efficient way to find the optimal MOF synthesis conditions represents the current bottleneck in speeding up MOF exploration. |
610936ae8804439fdde55e7b | 1 | synthesis is based on a time-consuming trial-and-error approach, in which a target MOF structure is compared with reported MOFs from literature to find similar synthesis conditions and experimentally refine them. A data driven approach (right loop), where a ML model is trained on a library of automatically extracted literature data, to then suggest synthesis conditions in a data-driven MOF discovery cycle. Updating the ML model based on new experiments leads to continuous improvement of the predictions. |
610936ae8804439fdde55e7b | 2 | The development of ML methods to predict the synthesis parameters for a desired MOF crystal structure based on scientific literature is a challenging but promising approach that will advance and accelerate chemical synthesis. Over the last years, ML methods have rapidly evolved, solving complex problems that involve highly nonlinear or massively combinatorial processes that conventional approaches fail to answer. Up till now, ML approaches have been successfully applied to address challenges in organic and inorganic synthesis. In the case of MOF synthesis, only recently, ML was used to optimize synthesis parameters for HKUST-1 and to determine the importance of the different parameters by analysing a set of partially failed experiments, in other words, "capture the chemical intuition" that can help to speed up the synthesis of similar MOF systems. However, the inverse synthesis design of MOFs, i.e. the automated prediction of suitable synthesis conditions for a targeted MOF structure (e.g. designed in silico) remains an unsolved challenge. This work represents a first step towards predicting synthesis conditions for an arbitrary MOF. We show a complete ML workflow for the inverse synthesis design of MOFs (going from crystal structure to synthesis conditions), (1) starting from automated data mining from scientific literature on MOF synthesis conditions and their structural information, (2) setting up and training of ML models, and (3) prediction of synthesis conditions for new MOF structures and comparison with human experts' predictions. |
610936ae8804439fdde55e7b | 3 | Data mining from MOF literature. To create a dataset with MOF synthesis parameters and structural information, we took advantage of the fact that well-curated MOF structural databases already exist (e.g. the Computation-Ready Experimental Metal-Organic Framework database CoREMOF and the Cambridge Structural Database CSD), in which MOF structural information and the corresponding publications with successful synthesis protocols are stored. The manual extraction of synthesis procedures from scientific literature is a time-consuming task, requiring the work of many experts. Alternatively, automatic data extraction to convert experimental procedures into a set of the desired synthesis parameters by employing natural language processing (NLP) techniques is a highly efficient and promising approach that we expect to be continuously improved in the upcoming years. In this study, we automatically extracted information on MOF synthesis for all publicly available MOF structures in the CoRE MOF database (SI Section 2.1). The six relevant parameters that were extracted are metal source, linker(s), solvent(s), additive, synthesis time, and temperature (Fig. ). To achieve this, we initially classified literature paragraphs, employing a decision tree with a string search method, to identify the synthesis paragraph related to each MOF structure (SI Section 2.2). After the synthesis paragraphs were determined, we employed the ChemicalTagger software, which focuses on the experimental part of a scientific text, recognizing significant words within the sentences, and annotating phrases inside the paragraph. In an effort to increase the tagging accuracy, we slightly modified the synthesis paragraphs, accounting for MOF-domain specific descriptions (SI Section 2.3). To evaluate the accuracy of the automatically extracted SynMOF-A database, we additionally generated manually corrected versions -the SynMOF-M and SynMOF-ME databases that are discussed in SI Section 2.4. |
610936ae8804439fdde55e7b | 4 | Alongside retrieving synthesis information from the MOF literature, we used the crystallographic information files (CIFs) from MOF databases to automatically extract the structural information of the linker and the oxidation state of the metal center. Ultimately, we combined the extracted synthesis details (i.e. metal source, linkers, temperature, synthesis time, solvents and additives) from the publications and information of the linker and the metal source from the CIF into the SynMOF database (Fig. ). Our central assumption in this work is that the established SynMOF database can be used for the training of ML models to facilitate the discovery of similarity patterns in the synthesis conditions to reach the final goal of predicting synthesis protocols for new MOF structures. Apart from the detailed information on MOF synthesis conditions, our SynMOF database, currently consisting of 983 MOF structures, provides the statistical data on the metal source and organic components (Figs. and). It contains 46 different metals with most common oxidation states ranging from +1 to +3. As expected, most MOF structures are composed of transition metals with copper and zinc comprising almost 50% of all metal types. Among the diverse organic molecules, the most commonly employed linkers for MOF synthesis are multidentate carboxylic acids (i.e. benzene-1,3,5-tricarboxylic acid, benzene-1,4-dicarboxylic acid, and benzene-1,2,4,5-tetracarboxylic acid) followed by N-containing bases (i.e. pyridine, triazole, and tetrazole). |
610936ae8804439fdde55e7b | 5 | In search of obvious patterns, we analysed the most common solvents used during MOF synthesis with respect to different temperature regimes and additives (Fig. ). At temperatures ranging from 80 °C to 160 °C, DMF and water, as well as their mixtures with other solvents are the most commonly used solvents. Synthesis at temperatures above 160 °C is predominantly carried out in water as a single solvent. Besides, the majority of MOF synthesis reactions at high temperatures (above 120 °C) are performed without additives, while at temperatures below 80 °C, the addition of acidic additives dominates. |
610936ae8804439fdde55e7b | 6 | SynMOF database, we trained multiple ML models to predict synthesis conditions of a diverse set of MOFs unseen during training. The input representation of the MOF structures is of crucial importance for the ML models performance. In this study, we used two types of representations as an input for the ML models training: One based on molecular fingerprints of the linkers, extended with encodings of the metal type and its oxidation state (Fig. , SI Section 3.1), and the recently developed MOF representation by Kulik and coworkers (SI Section 3.2). It is to be noted that the MOF field is still expanding, and an increasing amount of new structures and corresponding synthesis parameters will be available over time that can be used for training and refinement of ML models to achieve the highest possible performance. In this case, representation learning methods such as graph neural networks will then likely become more accurate than models relying on handcrafted feature representations. The prediction of synthesis time and temperature was achieved via regression models, such as random forests or neural networks (SI Sections 3.3, 3.4, 3.5). To predict discrete synthesis parameters, such as solvent and additives, classification models could be, in principle, used. However, for multiple reasons this turns out to be impractical: There is a wide variety of possible solvents and additives reported in literature, leading to a large number of categories, and, in turn, strongly imbalanced datasets. Furthermore, the properties of solvents can be very similar, making them interchangeable in synthesis, which leads to ambiguous solutions. In practice, also combinations of various solvents are required for successful MOF synthesis. Therefore, we developed a ML model which predicts solvent properties, such as partition coefficients, boiling point (SI Section 3.6), rather than the specific solvent. A nearest neighbor search in solvent property space yields lists of possible solvents that have properties similar to those predicted by the ML model. In this way, new solvents can be incorporated easily, and even solvents occurring only once in literature can be used to train the model. In the case of additives, we found that the main parameter that distinguishes different additives is their acidity/basicity strength. Thus, we split the dataset into three groups (acidic, basic or no additive) and used a classification model for additive prediction. |
610936ae8804439fdde55e7b | 7 | The results of our trained ML models are shown in Fig. . Reproducibly positive correlation coefficients r 2 on unseen test datasets show that the ML models are capable of identifying meaningful and predictive relations between the target MOF structure and the required synthesis conditions, in particular temperature and time (Figs. ). Given the amount of data that we have currently extracted from literature, we find that the random forest models have the highest performance across all predicted parameters. However, neural networks learn to make better predictions with growing dataset sizes faster (see learning curves in Fig. ) and even exploit correlations between different synthesis parameters (e.g., solvent and temperature) rather than predicting them separately. Hence, we expect that more complex models will outperform random forests in the near future. |
610936ae8804439fdde55e7b | 8 | To evaluate ML-based solvent prediction, we focused on a subset of MOFs which are synthesized using only one solvent. We compared the accuracy of the top 6 ML predictions with multiple random baseline methods (Fig. ), including selection of a random solvent out of all solvents as well as out of the six most frequent solvents that are used in 96% of the single-solvent SynMOF database. We found that the ML model outperforms the random selection, in particular for the top 1 -3 solvent predictions, where the ML model reaches an accuracy of > 90%. In the case of additive predictions (Fig. ), the task of the ML model is to classify required additives as acidic, basic, and no additive. While performing well on the training set, the generalization to unseen test data suffers from an imbalanced dataset (most database entries do not use an additive). We use balance correcting weights of the training data points, leading to predictions which distinguish very well between synthesis procedures involving basic and acidic additives. However, the differentiation between acidic and no additive or basic and no additive is less pronounced. One of the reasons might be related to the hidden variables such as type and function of additives: Some of them (inorganic acids and bases) have only the role of pH regulation, while others (organic acids and bases) are also involved in modulation of the MOF growth. Besides, concentration and strength of additives are additional important parameters, influencing the role of additive. |
610936ae8804439fdde55e7b | 9 | To put the ML performance into perspective, we performed tests with 11 human MOF synthesis experts. We developed an online quiz based on 50 MOFs randomly selected from the SynMOF database which will be publicly available. The participants were given the 3D structures of MOFs, chemical structures of the linkers and information on the metal ion, and asked to estimate synthesis conditions such as temperature, time, solvents and additives without any help from literature or other external sources (SI Section 3.7). After each MOF synthesis prediction, we also asked the participants to estimate how certain they are in the answer. The correlation coefficients r 2 between the experts' temperature and time predictions and the reported synthesis conditions are close to zero, even after averaging over 11 estimates by different researchers (Fig. ) and after sorting only by predictions with high certainty. This rather surprising result shows that even small correlations learned and exploited by the ML model will help to estimate better synthesis conditions. In summary, we showed that the ML models are able to learn generalized patterns and correlations in the SynMOF database, which exceed the experts' general intuition, and thus, could be used to identify good first guesses for experimental synthesis attempts of new MOFs. |
610936ae8804439fdde55e7b | 10 | The lack of machine readable and curated MOF-synthesis data up till now hindered the development of digital ML tools for predicting MOF synthesis conditions. Here, we established a SynMOF database by automatic data extraction via NLP methods that provides synthesis conditions and structural information for more than 900 MOFs, and trained ML models based on these data to identify patterns in MOF synthesis. We expect that the created SynMOF database will boost the NLP research within the MOF community, while our ML synthesis prediction platform will be the new gold standard for data-driven MOF discovery. Even at an initial stage, our ML models outperformed MOF experts' synthesis prediction, underlying both the complexity behind the synthesis process and a pressing need in developing digital predictive tools. Our automated on-demand synthesis prediction will considerably accelerate the discovery of new MOFs and serve a valuable tool for the MOF community and beyond. |
Subsets and Splits