id
stringlengths 24
24
| idx
int64 0
402
| paragraph
stringlengths 106
17.2k
|
|---|---|---|
62a3951999b7836e25fbe614 | 23 | In summary, we have introduced a workflow that trains high quality NN to learn the underlying physics of inter-atomic and intermolecular interactions from existing classical poten-tial models and transfer the learning to a network that can be retrained with higher fidelity DFT dataset. Such an approach is shown to overcome the limitations of pre-defined functional forms employed in classical MD -our approach builds on the physical models and incorporates flexibility as well as accuracy via retraining with a DFT dataset. The transfer learning workflow also includes an active learning loop that generates a NN potential that does an excellent job of replicating DFT energies over an elaborate energy range for 10 different elemental systems, while using sparse training data sets. Our method first learns the radial contributions to the potential energy surface of the 10 elements using the SC EAM potentials followed by transfer learning of the network using DFT as the reference that also include the angular contributions. |
62a3951999b7836e25fbe614 | 24 | Our approach overcomes a significant drawback of SC EAM potentials which works well for the bulk systems but fails in capturing the energetics of the nanoclusters. With a relatively sparse dataset on the order of a few hundred SC EAM augmented with DFT calculations, we build a NN model that is able to concurrently capture the cluster as well as bulk properties over an extensively sampled DFT test set of 2000 -3000 configurations for each of the 10 elements. We show that our NN provides excellent predictions of both the energies and forces over a wide variety of cluster sizes. Moreover, it also reasonably captures the structure, cohesive energies, elastic and thermodynamic properties of bulk condensed phases, in good agreement with DFT calculations. |
62a3951999b7836e25fbe614 | 25 | The Vienna Ab-initio Software Package (VASP) [1] along with the Perdew-Burke-Eznerhof (PBE) [2] exchangecorrelation functional was used to perform all density functional theory (DFT) calculations for calculating energy and forces of clusters in the training and validation datasets. The projector-augmented wave (PAW) potentials used in these calculations are summarised in Table . A single k-point at the center of the Brillouin zone (gamma point) was used for each calculation. The convergence criteria for the electronic self-consistent iteration and the ionic relaxation loop were set to be 0.1 meV and 1 meV per cluster, respectively. The ground state bulk structures for the elements were collected from the Materials Project database [3]. A dense k-point mesh, (n atoms × n kpoints ≈ 1000) was used for the DFT calculations that were used to calculate the cohesive energies and lattice parameters. A relatively high tolerance of 10 -6 eV for energy convergence was employed for the same. |
62a3951999b7836e25fbe614 | 26 | We evaluated the performance of several common empirical force field models [4, 5] that are regularly adopted for Molecular Dynamics simulations. We compared the energy and force predictions of well-known and a wide variety of potentials (including EAM [6, 7], MEAM [8, 9, 10], ADP [11, 12]) against that of the DFT values. Comparison results across 10 elements, covering over 11 different potentials are described in this section. These results are used as a benchmarking starting point for our ANN force field Models rewrite the last sentence. |
62a3951999b7836e25fbe614 | 27 | Table : Summary of the force fields collected from literature to predict nanocluster energies and forces. The references for each force field is mentioned in the square brackets. [13] Element Functional Form Ag eam/fs [14], eam/alloy [15], sutton chen [16] Al eam/fs [17], meam [18], sutton chen [16], adp [19] Au eam/fs [14], eam/alloy [4], adp [20], sutton chen [ The ANN Forcefield models are validated against a set of 30000 structures for each element created using nested sampling, with energies calculated using SC-EAM. Both Force as well as energy calculations are conducted and the correlation plots for the optimised final NN from activelearning are shown in this section. The training sets for these networks are used in the next phase to retrain the NN using DFT The NN trained using SC-EAM is validated against a validation structure set calculated using DFT to serve as a benchmark for the ANN retrained using DFT. The lack of angular components in the fingerprints used for these networks makes it difficult for them to accurately capture the energy and force data calculated using DFT. The Retrained ANN Forcefield models are validated against a set of 1600-2000 Structures whose energy and forces were calculated using VASP. The mean absolute errors for energies(in meV/atom) and forces (in meV/Å) are included in the legend of the plot. Each plot has four panels; top row containing energy and Force plots for the training, while the bottom row covering energy and Force plots for the test data. |
6523baabbda59ceb9a29e511 | 0 | Additives are chemical compounds added to polymer matrices, in relatively low concentrations, except for some as fillers/reinforcements or plasticizers. Without additives, polymers would have virtually no properties for use, and often too little resistance over time. Additives play a crucial role in the behavior of materials in terms of appearance and properties . They can also influence durability, stability, UV resistance, fire resistance... Additives can be classified into 4 distinct categories : (i) functional additives (antioxidants, plasticizers, anti-UV…), (ii) colorants, (iii) fillers and (iv) reinforcements. Several literature reviews have already been published in this field , which shows that the choice of an additive is based on numerous criteria, such as the nature and structure of the macromolecules. The additive must be compatible and dispersible into the matrix. |
6523baabbda59ceb9a29e511 | 1 | Plasticizers are among the most widely used additives in the industry, with a market estimated at $17 million, representing over 60 % of the additives market . They are defined by IUPAC as « a material incorporated in a polymer to increase the polymer's workability, flexibility or extensibility » . They are essential for enhancing polymer flexibility and mobility, improving processability , with a great reduction of the glass transition temperature (Tg) of the matrix. Plasticizers also decrease the stiffness of the matrix as represented by Young's modulus (E) and increase the elongation at break (εmax) . They modify intermolecular forces and polymer crystallinity, by increasing the free volume and the chains mobility . They thus make it possible to adjust the thermomechanical properties of a material for a specific application or process. For example, some thermoplastics are too rigid and brittle to be easily shaped by conventional thermomechanical processes. Thanks to plasticizers, these polymers could be shaped more easily, acquiring new properties, opening up new fields of application. |
6523baabbda59ceb9a29e511 | 2 | Polyvinyl chloride (PVC) is the most plasticized of all polymers, giving it new properties and applications in e.g., packaging, construction materials, toys, healthcare... Over 80 % of plasticizers produced are used for PVC . Phthalates, which are esters of phthalic acids, represent the most common PVC external plasticizers, such as bis(2-propylheptyl) phthalate (DPHP) or diisodecyl phthalate (DIDP), whose structures are shown in Figure . These phthalates have many advantages, such as excellent compatibility, low volatility and low cost . The carbonyl groups are able to form weak bonds with PVC on the C-Cl dipole and thus increase their compatibility and miscibility, forming a network based on weak interactions. In addition, the fatty chains from these plasticizers increase the inter-macromolecular free space between the chains and thus bring mobility. Nevertheless, these external plasticizers can partially migrate outwards by exudation and/or volatilization, leading to problems of phthalate bioaccumulation in the human body and the environment . These fossil-based additives are suspected of causing health problems, such as endocrine disruption. Regulations have long been in place to limit the use of phthalates notably in toys for young children . Then, since 1999, the European Commission has banned the use of phthalates in children's toys. In 2005, an EU directive (2005/84/EC) made this ban permanent. This directive also extended the ban to childcare articles . |
6523baabbda59ceb9a29e511 | 3 | As a result, the interest in new plasticizers of renewable and sustainable origin has risen rapidly in recent years, as shown by an abundant recent literature. For instance, a multitude of biobased esters have been tested. Several literature reviews present plasticizers based on esters of vegetable oils, citric acid (called citrates) or sugar derivatives such as isosorbide . |
6523baabbda59ceb9a29e511 | 4 | Poly(hydroxyalkanoate) (PHA) are non-toxic (co)polyesters, which are bioproduced by a wide range of bacteria as intracellular storage materials. These bacterial polymers can be synthesized by fermentation from a wide variety of carbon-based substrates. Different biomasses or waste plastics can be used as carbon sources. For example, poly(ethylene terephthalate) (PET) can be depolymerized at the end of its life by enzymatic or chemical hydrolysis into terephthalic acid and ethylene glycol. |
6523baabbda59ceb9a29e511 | 5 | These monomers can then be converted by genetically modified bacteria to produce PHAs . Depending on the substrate structure and the enzymes involved, a wide range of bacterial polymers can be obtained, differentiated by their number of carbons per repeat unit . Among them, medium side-chain length (mcl) PHAs have repeat units of 6 to 12 carbons, such as poly(3-hydroxy-octanoate) (PHO). The chemical structure of PHO is shown in Figure . |
6523baabbda59ceb9a29e511 | 6 | In this study, PHO is used as a plasticizer precursor thanks to the presence of (i) ester groups and (ii) relatively long pendant alkyl chains. Various plasticizers are synthesized from PHO and evaluated with PVC and rigid TPU matrices in a preliminary and conceptual approach. In a first step, the initial molar mass of PHO was reduced in a controlled manner to improve its plasticizing effect . |
6523baabbda59ceb9a29e511 | 7 | Hydroxyl-terminated oligomers (PHO-diol) were obtained by transesterification when BDO was selected as the diol. This latter can notably be obtained from biomass or from waste plastics, such as poly(butylene adipate terephthalate) (PBAT) or poly(butylene succinate) (PBS), which are depolymerized by chemical hydrolysis . The chain ends of PHO-diol were then modified with hexyl (mono)isocyanate (HI), to obtain an end-capping with short alkyl chains (PHO-hexyl) with the aim to improve the additive's dispersion into the matrix, and its plasticizing effect. To test the different additives, formulations were made from rigid PVC and TPU, at different plasticizer concentrations. Commercial references based on phthalate groups, such as DPHP and DIDP were used. The chosen TPU was synthesized with a high content in cycloaliphatic sustainable betulin, which brings high rigidity and brittleness to the polymer, preventing its processability and thus making it hardly exploitable. These different formulations were then evaluated and characterized to determine the plasticizers efficiency, initially in terms of Tg evolution and mechanical properties. |
6523baabbda59ceb9a29e511 | 8 | Poly(tetrahydrofuran) (PTHF) with a weight-average molar mass (Mw) of 650 g mol -1 (PTHF 650) was acquired from BASF (Germany), under the brand name PolyTHF®. Betulin (98%) was purchased from NST-Chemicals (USA). PHO was kindly bioproduced and supplied by the University of Dublin in the frame of MIX-UP Project (Horizon 2020 research and innovation program under grant agreement No 870294). PHO was bioproduced according to the protocol given by Cerrone et al. . |
6523baabbda59ceb9a29e511 | 9 | PHO-diol was prepared from native PHO by transesterification reaction with BDO. The synthetic route for this oligomer is shown in Figure , according to a previous protocol . In a two-neck round-bottom flask with a water cooler, 60 g of PHO are dissolved in diglyme (250 mL) and BDO (150 eq/OH), then magnetically stirred at 100 °C under argon until completely dissolved. The temperature is then adjusted to 160 °C. Then, DBTL (0.5 eq/OH) is added, followed by stirring under argon for 48 h. The progress of the reaction is monitored by 1 H NMR until a number-average molar mass (Mn) of around 800 g mol -1 is obtained. The reaction is then cooled down before purification. The product is dissolved in chloroform and washed three times with distilled water to remove the BDO in excess. The product is then distilled to remove traces of diglyme and water. An orange viscous oil is obtained. |
6523baabbda59ceb9a29e511 | 10 | Quantitative 31 P NMR samples were prepared by dissolving 15-20 mg of product in 500 µL of CDCl3, following a protocol from the literature . Next, 100 µL of a standard cholesterol solution (0.1 mol L -1 in a 1/1.6 CDCl3/pyridine mixture) containing Cr(acac)3, an NMR relaxation agent, was added along with 50 µL of 2 chloro-4,4,5,5-tetramethyl-1,3,2-dioxaphospholane. The number-average molar mass (Mn) of PHO-diol oligomers was obtained from 1 H NMR using Equation . 𝐼 3,6 represents the intensity of the peak at δ = 3.67 ppm, corresponding to the two protons in position α of the primary hydroxyl groups at the end of the chain. The end-of-chain unit with these chemical groups has a molar mass of 89.11 g mol -1 . 𝐼 2,4-2,2 represents the intensity of the peak between δ = 2,4 and 2,2 ppm which is based on the protons in α-position of the secondary OH groups at the end of the chain. The end-of-chain unit possessing these groups has a molar mass of 143.21 g mol -1 . 𝐼 2,6-2,4 represents the intensity of the peaks between δ = 2.6 and 2.4 ppm of the CH 2 in position α of the ester, corresponding to a hydroxy octanoate (HO) repeating unit with a molar mass of 142.20 g mol -1 . |
6523baabbda59ceb9a29e511 | 11 | Quantitative 31 P NMR is used to determine the hydroxyl value (iOH in mg KOH g -1 ) of PHO-diol for the PHO-hexyl synthesis step, using Equation , where m is the sample mass in g. C and V are the concentration (in mol L -1 ) and the volume introduced (mL) of the standard solution, respectively. 𝐼 1 is the intensity of peaks between δ = 133.5 and 134.6 ppm, corresponding to primary alcohols at the end of the chain. 𝐼 2 is the intensity of peaks between δ = 135.0 and 137.0 ppm, corresponding to terminal secondary alcohols. 𝐼 𝐵𝐷𝑂 is the intensity of the peak at δ = 134.2 ppm, corresponding to residual unreacted BDO. 56.1 is the molar mass of potassium hydroxide (KOH) in g mol -1 . |
6523baabbda59ceb9a29e511 | 12 | Different formulations with PVC and TPU matrices were prepared with the addition of plasticizers at different content (wt%), within plausible ranges for an application context (from 15 wt%). According to the literature, plasticizers contents between 15 and 50 wt% are recommended for PVC plasticization . Moreover, the concentrations chosen are well beyond a potential antiplasticization range commonly found at low concentrations for most plasticizers especially in PVC . The formulations are presented in Table . To achieve this, 15 g of PVC are added to 150 ml THF at 50 °C and stirred until completely dissolved. |
6523baabbda59ceb9a29e511 | 13 | The appropriate amount of plasticizer is then added. For the commercial plasticizers DPHP and DIDP, a 50/50 (wt%) mixture is made, denoted "DPHP:DIDP". The solution is stirred for a further 10 min, then poured into a glass Petri dish and left overnight in a fume hood for solvent evaporation. The final material was then placed in a vacuum oven at 50 °C for 6 h to remove the traces of solvent. |
6523baabbda59ceb9a29e511 | 14 | 10 g of a sustainable TPU was synthesized in a conventional two-step process with MDI as the diisocyanate, PTHF 650 as the polyether polyol and betulin as the chain extender. A prepolymer NCO/OH number (iNCO) of 3 and a final iNCO = 1 were chosen to obtain a specific rigid polymer with high Tg. The synthesis route is shown in Figure . After overnight to finalize the reaction in a vacuum oven at 50 °C, 100 mg of TPU were dissolved into 5 ml THF at 60 °C. Then, an appropriate amount of plasticizer is added, and the resulting solution is stirred for 5 min. The mixture is poured into a glass Petri dish and left overnight in a fume hood to evaporate the solvent. Traces of solvent are then removed in a vacuum oven at 50 °C for 6 h. |
6523baabbda59ceb9a29e511 | 15 | Thermal analysis by differential scanning calorimetry (DSC) was carried out using a TA Instruments Discovery DSC 25. Approximately 10 mg of polymers were placed in sealed aluminum capsules, then analyzed under nitrogen flow (50 mL min -1 ). A three-step procedure with ramps of 10 °C min -1 was applied as follows: (i) heating from 90 °C to 200 °C to erase the thermal history of the samples, (ii) cooling to 90 °C and (iii) heating from 90 °C to 200 °C. Tg values are determined during the second heating step. |
6523baabbda59ceb9a29e511 | 16 | Experiments were carried out at 20 °C with a constant crosshead speed of 20 mm min -1 . Series of at least five samples with dimensions of around 30 × 10 × 1 mm 3 were taken. Young's modulus (E), strength at break (σmax) and elongation at break (εmax) were calculated. To produce the specimens, 1 mm-thick films were hot-pressed using a LabTech Engineering Company Ltd. press. For PVC, around 13 g of polymer were first preheated to 150 °C for 3 min in a 10 x 10 cm 2 mold, and then pressed between the plates at 160 bar and then cool down till room temperature for 5 min. |
6523baabbda59ceb9a29e511 | 17 | However, its Mn is 21.6 kg mol -1 , which is too high for a plasticizer . A transesterification reaction strategy with a short diol was thus considered to obtain controlled low Mn oligomers. As PHO is not soluble in BDO, a sustainable reactive solvent, diglyme was therefore used as a co-solvent with high-boiling point, for a reaction at 160 °C. The reaction is shown in Figure . The progress of the reaction is monitored by 1 H NMR until to obtain a Mn close to 800 g mol -1 . After 48 h of reaction, a plateau is reached, and the Mn no longer varies. The Mn of the starting PHO and the final PHO-diol are measured by 1 H NMR and SEC and reported in Table . Mn value is thus reduced by 18 during the reaction. This Mn can be reached in 1.5 hours by increasing the amount of DBTL to 50 eq/OH. |
6523baabbda59ceb9a29e511 | 18 | Indeed, the OH groups at the end of the PHO-diol can (i) improve plasticization by creating interactions between the plasticizer and the matrix, but also (ii) limit a good dispersion in the matrix by locally creating weak bonds with other plasticizer molecules to form micro-aggregates. To this end, studies have shown that grafting long alkyl chains onto polar chain ends is a good strategy for adapting the ratio between the number of polar and apolar groups and obtaining good chain mobility. |
6523baabbda59ceb9a29e511 | 19 | Then, a second synthesis step was carried out to add alkyl chain ends that could both promote additive dispersion in the matrix, but also increase the free volume between polymer chains and thus their mobility. To this end, a fast conversion reaction of PHO-diol to PHO-hexyl was carried out with HI under stoichiometric, solvent-free conditions. The structure of PHO-hexyl was confirmed by 1 H NMR (SI, Figure ). |
6523baabbda59ceb9a29e511 | 20 | The greater the variation in Tg compared with the matrix alone, the greater the plasticizing effect with an addition of mobility. The Tg values are plotted in Figure . The addition of plasticizer leads to a very significant decrease in Tg, between 14 and 64 °C, for the different formulations. This result is the consequence of various associated phenomena which are linked to (i) the strong C-Cl dipole of PVC which interacts with the different dipoles carried by the plasticizers (ester and urethane functions) creating weak stabilizing interactions (limiting exudation) and forming a relatively structured network, and (ii) the pendant chains which themselves bring free volume and therefore mobility to the network formed. |
6523baabbda59ceb9a29e511 | 21 | The Tg decrease is more pronounced for formulations based on PHO-hexyl than on PHO-diol ones. It can also be noted that formulations containing PHO-hexyl have Tg comparable to those based on the conventional and commercial plasticizer mixture DPHP:DIDP. Indeed, the Tg decreases by only 20 °C for 45 wt% of PHO-diol (Tg = 63 °C), whereas it decreases by 50 °C for 30 wt% PHO-hexyl (Tg = 32 °C). |
6523baabbda59ceb9a29e511 | 22 | Same trends have been recently observed in the literature. A decreased of Tg of 77 °C was obtained with 30 wt% aliphatic polyesters with unprotected chain ends with PVC . A recent study showed that it is possible to adjust the plasticizing effect and therefore the Tg of materials, by playing on the length of the alkyl chain grafted onto polyester chain ends . For example, Tg values below 30 °C have been achieved with alkyl chain ends with 8 or 12-carbons. Indeed, the OH groups at the end of the PHO-diol chain may tend to form aggregates due to the hydrogen bonds formed between plasticizer molecules. As a result, PHO-diol may be poorly distributed in PVC, reducing the efficiency of the plasticizing effect. Thus, the alkyl chain ends of PHO-hexyl would enable better dispersion in the matrix, beyond the increase in free volume. Uniaxial tensile tests were also carried out to determine the evolution of E and εmax of formulations containing PHO-hexyl and DPHP:DIDP. The results are reported in Figure . All values for plasticized PVCs are similar, indicating a similar plasticizing effect between phthalates and the sustainable additive. Moreover, σmax and E were at least divided by 6 and 13, respectively, compared to neat PVC values. The plasticization effect reduces the rigidity of materials and thus lowers E and thus σmax. As a result, σmax increased by a factor of at least 12, confirming the plasticizing effect of these molecules. |
6523baabbda59ceb9a29e511 | 23 | The field of TPU plasticization is marginal compared to that of PVC one. To analyze the potential of the plasticization of TPU, a particular renewable rigid TPU, hardly exploitable in its current state because of low processability, was synthesized and then plasticized. The objective is to show the interest and the contribution of the plasticization on a TPU that is hardly exploitable in its current state, as the worst possible case. |
6523baabbda59ceb9a29e511 | 24 | This highly rigid and sustainable TPU was synthesized using a conventional 2-step protocol from MDI, PTHF 650 (a medium-long polyether diol) and betulin (a rigid chain extender). The synthesis route is shown in Figure . Betulin is a cycloaliphatic terpene which can be obtained by chemical and/or biotechnological synthesis from sustainable resources such as plastic waste or biomass. The TPU rigidity is mainly provided by MDI and its aromatic rings and the betulin, which features 5 aliphatic rings. The obtained TPU is very brittle and has a high Tg (51 °C) as determined by DSC (Figure ). PHOhexyl and DPHP:DIDP were incorporated into TPU matrix as plasticizers at different content (Table ) |
6523baabbda59ceb9a29e511 | 25 | In rigid TPUs, the free volume is low due to numerous hydrogen bonds between the various rigid segments linked to the urethane groups. PHO-hexyl's action as a plasticizer is therefore multiple, due to its chemical structure. PHO-hexyl's ester and urethane groups can form interactions with TPU urethane groups, in particular. TPU-plasticizer interactions thus reduce TPU's own intra-and interchain interactions. They also minimize potential phase separation (demixing) with plasticizer volatilization or migration towards the sample surface. The alkyl-pendant chains of the HO units allow free volume to be gained by spreading the rigid TPU chains apart, which also reduces interactions between them. PHO-hexyl therefore has also a great potential as a plasticizer for rigid TPUs, even in small content. This neat rigid TPU could not be directly shaped by thermo-compression, as it was far too stiff and brittle to be demolded (Figure ). After plasticizing with 15% DPHP:DIDP or PHO-hexyl, the TPU-based materials could be processed and shaped into films. This clearly demonstrates the value of plasticizing rigid TPUs, opening up a whole new field of applications. The plasticized materials were then characterized in terms of mechanical properties using uniaxial tensile tests. The results are shown in Figure . The εmax of the two plasticized TPUs are similar, while the E and σmax are about twice as high for the TPU plasticized with PHO-hexyl. |
6523baabbda59ceb9a29e511 | 26 | Phthalates and PHO-hexyl show a plasticizing effect improving the processability and increasing TPU flexibility, with better results for DPHP:DIDP. This also shows that a small gap (some °C) on the Tg of a plasticized polymer (Figure ), can have a much greater impact on the mechanical properties and processability of the final materials, as also shown in the literature . Both rigid PVC and TPU matrices were used to evaluate the different plasticizers at various contents: |
6523baabbda59ceb9a29e511 | 27 | PHO-diol, PHO-hexyl and two conventional phthalates (DPHP and DIDP). The Tg of the polymers were significantly reduced, giving the material greater flexibility. Plasticizing effects similar to those of classically used phthalates were obtained with PHO-hexyl. The alkyl chain ends of PHO-hexyl enabled better dispersion of the additives in the matrices, as well as an increase in free volume, which was reflected in further improvement of the thermal and mechanical parameters compared to PHO-diol. |
6523baabbda59ceb9a29e511 | 28 | PHO-hexyl thus appears to be a very promising example of a sustainable plasticizer, and as a potential credible alternative to phthalates in this preliminary and conceptual study. This new plasticizer appears to offer the potential to produce high-performance, environmentally friendly materials. It is possible to adjust the final properties of the materials, by adjusting the amount of plasticizer and the size of the grafted chain ends. |
6523baabbda59ceb9a29e511 | 29 | The prospects for this work and the potential additions to it are very broad. There are over 150 different chemical structures of PHAs, depending on the carbon substrate used and the biosynthesis conditions. It would therefore be possible to develop a very wide range of sustainable plasticizers, using a "chem-biotech" approach. This study needs to be extended in a number of ways, in particular to assess the mechanical properties and rheological behavior of the systems developed. It is crucial to consider plasticizer migration studies as an obvious perspective to this first conceptual approach for the valorization of PHO-based plasticizers. Moreover, the hexyl isocyanate used to add alkyl chain ends is highly toxic, thus another synthesis route could be developed to reduce its toxicity. In this frame, it would also be necessary to establish toxicity tests for the systems developed for a plausible approach to replacing phthalates for a future that is more respectful of the environment and health. |
64d660394a3f7d0c0d01f530 | 0 | The domain of π-conjugated polymer semiconductors is of keen interest to both the materials informatics and organic electronics communities due to the promising opportunities this class of materials offers for large-area, printable, deformable electronic devices and energy applications. In the years since the advent of the Materials Genome Initiative in 2011, conjugated polymer semiconductors have been associated with over 5,000 peer-reviewed articles (Web of Science, March 2023). A subset of this body of literature has striven to further accelerate knowledge discovery in this materials space through applied data science, machine learning, and high-throughput experimentation techniques. For example, data-driven techniques have recently been leveraged to pursue targeted advances for thin-film device applications including organic field-effect transistors (OFETs), 2, 5-8 organic light-emitting diodes (OLEDs), and organic photovoltaics (OPVs). Early successes include the application of self-driven laboratory workflows to screen quaternary OPV formulations at the full device level. Indeed, these endeavors have positioned organic electronics as a significant research thrust within the materials informatics community. |
64d660394a3f7d0c0d01f530 | 1 | Despite recent accomplishments, rational design of organic electronic devices, particularly those that are polymer-based, still largely materialize through one-parameter-at-a-time, hypothesis-driven studies due to the limited availability of representative experimental data. The conjugated polymer materials domain is a research area with a compelling need for experimental data management solutions. While a few examples of shared datasets or databases that target organic electronics research have been reported, such as the Harvard Clean Energy Project and OCELOT, they mostly include computational data on small molecules. A recent effort in Deep4Chem mined over 1,000 peer-reviewed articles to build an experimental database of chromophores, but similar to prior efforts it largely targets electronic structure-property measurements and is not inclusive of process history. In extending database management effectively to polymer-based devices, providing data models that are inclusive of experimental processing information are priorities for storing accurate and reproducible records. |
64d660394a3f7d0c0d01f530 | 2 | Experimental database design and management for polymer electronics however is nontrivial, especially when a plurality of the relevant attributes related to the provenance of the sample must be included accurately to form a robust, "reusable" data record. Data ontologies are not standardized in the organic electronics space: fully capturing all relevant experimental information is challenging, and organic device performance is highly sensitive to the many parameters associated with the active layer deposition process. For example, the charge-carrier mobility (μ), a key figure of merit for OFETs, has been shown to vary significantly for poly(3hexylthiophene) (P3HT) (~10 -6 -10 0 cm 2 /V•s) cm 2 /V•s). This performance variation is not only attributed to batch-to-batch characteristics of the polymer, but also a plethora of parameters related to the polymer's process history, starting with the solution state through the thin-film deposition process. Another source of variation is that mobility values are derived from device measurements via model fitting, and employing different methods/parameters (i.e., models, measurement settings, voltage limits) may affect the extracted mobility value. Recording processing and measurement parameters provides indispensable contextual value to organic device data, but nonetheless recording all of them efficiently is not straightforward. Additionally, since the design space is inherently dynamic due to the evolving nature of research, data models must be designed with flexibility in mind without sacrificing consistent vocabulary. Thus, generating a representative data ontology describing the experimental device realm is a challenge that must be addressed to enable reliable database designs. Though process representations are not new problems for the sake of curating materials databases, navigating these challenges for experimental polymer domains has only been explored recently by a minority of materials data researchers. The experimental database effort in MaterialsMine promotes the inclusion of processing terms for polymer nanocomposites, while the Community Resource for Innovation in Polymer Technology (CRIPT) proposes a framework to comprehensively describe polymer data, seeking to unify all aspects of sample provenance from synthesis, processing, characterization, properties, and instrumentation/citation metadata. These active endeavors open the door for a broader adoption of polymer-based data management solutions, but it is up to various communities to enable tailored data models for their specific subdomains. In this work, we use OFETs as a model system to propose an experimental data ontology associated with semiconducting polymer processing. We then produced a data structure that focuses on the deposition of the active semiconducting polymer layer and leveraged it to implement an experimental repository relating the semiconducting polymer process history to device performance. To guide a robust representation of that process history, this work draws upon ISA-88, an international standard for automation in batch process control, to construct generalizable relationships across process transformations within the fabrication procedure to create the semiconducting thin film. Building a data repository that can handle the many nuances of this complex design space is expected to provide a platform to enhance hypothesis design, scientific decision-making, and model development within the traditionally "small data" space of the organic electronics community. |
64d660394a3f7d0c0d01f530 | 3 | Defining the required information to capture is facilitated by published reporting standards for experimental OFET device data. An overview of the major materials and process stages involved in depositing the semiconducting polymer layer, and a non-exhaustive set of their related parameters/attributes is presented in Figure ! Reference source not found.. A device recipe considers the starting materials -a polymer and a device substrate -and tracks these two inputs as they are transformed through a series of process steps and ultimately integrated into the output: an OFET on which a device measurement is made. Important parameters and nuances to evaluate device performance include materials characteristics, solution processing, substrate treatment, and the instrument parameters and models used to extract device metrics. |
64d660394a3f7d0c0d01f530 | 4 | The primary material parameters describe the components of the active semiconducting layer as well as any other materials (i.e., solvents, chemical treatments) included in the processing procedure. The polymer and any other components are dissolved into a solution that is ultimately deposited onto the device substrate. As the behavior of the polymer in the solution state is a key determinant of its thin-film behavior, all information associated with the solution makeup and its processing is especially crucial to capture for data provenance purposes. Choices in the material characteristics of the polymer (i.e., molecular weight distributions, regioregularity, tacticity, etc.) yield a range of structural and morphological motifs that in turn significantly impact device performance. The identity of the solvent(s) used affects not only the polymer-solvent interactions in the solution-state, but also the dynamics of the thin-film deposition process, thereby influencing the structure and performance of the final active layer. The OFET device substrate is a layered device structure, comprising substrate, gate electrode, dielectric, and source/drain electrodes. The most common electrode configurations used in the organic electronics community include (a) bottom-gate bottom-contact (BGBC), (b) bottomgate top-contact (BGTC), (c) top-gate bottom-contact (TGBC), and (d) top-gate top-contact (TGTC). The electrode configuration is a necessary contextual detail for reporting a device measurement, as substrate designs often take advantage of charge carrier behavior at different interfaces. Additionally, as substrate design parameters (i.e., channel dimensions such as width and length) and material choices (i.e., electrode material, dielectric, etc.) may influence device measurements, this design information is important to include in a data entry to promote experimental reproducibility. An emphasis of the work herein is that processing information is indispensable for the purpose of storing reproducible device data. Seemingly minor differences in processing can lead to significant changes in the recorded charge-carrier mobility of an OFET sample. Omitting information related to this process history will therefore lead to errors or inaccuracies when comparing device data across experiments. Even prior to deposition, the solution and device substrate undergo process transformations that can affect the deposited thin-film and device characteristics. Solution-based processes may include operations such as sonication, aging, poor solvent addition, cooling, etc. 39 in a prescribed sequence to promote solution-state aggregation, while surface pretreatment procedure may include, for example, a cleaning process (e.g., UVozone or plasma treatment) followed by a surface modification step via self-assembled monolayer (SAM). As an example, differences in solution aging times (i.e., 3 hr, 6 hr, 24 hr prior to coating) can lead to noticeable structural changes that affect the final value of the device measurement in P3HT. The coating process could be performed through a plethora of solution casting methods including drop casting, spin coating, blade coating, inkjet printing, slot die coating, etc. that all have different physical impacts on thin-film morphology and therefore device performance, especially when coupled with solution pre-treatment. Meniscus-guided coating techniques, for example, yield a set of deposition regimes governed by a complex parameter space that includes flow conditions, coating speeds, stage temperatures, drying times, contact angles, etc. In combination with solution properties and surface interactions, coating conditions are often chosen carefully to tune the morphology of the deposited thin film. Post-processing operations may also be performed after the coating stage, such as annealing or selective etching, to further control the thin-film morphology. |
64d660394a3f7d0c0d01f530 | 5 | Particularly, OFET performance metrics can be nontrivial to represent because a charge-carrier mobility value is not a measurement per se; it is a parameter value derived from curve fitting of the actual measurement, a transfer curve sweep. Unreported details about measurement regimes, voltage sweep range/direction, the measurement environment, etc., can lead to misinterpretations about the provenance of device metrics such as the mobility. The importance of measurement metadata reporting applies across all types of measurements, as elaborated on elsewhere. Inputs to the process recipe are represented by material nodes and process stages (see Figure ). |
64d660394a3f7d0c0d01f530 | 6 | At a high level, the entity-relationship diagram in Figure shows how an organic device sample with its associated reported measurement (i.e., a charge-carrier mobility value) may be conceptually encapsulated as an experimental data record. This diagram provides an important visualization of how various parameters, data, and information in the experimental real world are captured as attributes of related objects, to facilitate the organization of data in constructing a database. Rectangles represent entities or objects, diamonds represent relationships between entities, and labeled ovals represent attributes containing the data or information associated with the various entities, where underlined labels denote a unique identifier for that object. |
64d660394a3f7d0c0d01f530 | 7 | Sample has a one-to-many (1-N) relationship with measurement, since a single OFET could be measured a number of times by a variety of characterization methods not limited to device performance. A measurement has a type (e.g., transfer curve, spectroscopy, scattering, thickness, etc.) and the heterogeneous data (e.g., value type, value, error, etc.) and metadata (e.g., instrument information, date measured, etc.) associated with it. An experiment refers to the source information associated with the reporting of the sample, and therefore has a citation type (e.g., laboratory, journal article, dissertation) and metadata associated with that source (e.g., digital object identifier (DOI), date published, author information, etc.). Experiment and sample have a one-to-many relationship, as a single experiment may report multiple samples. A process recipe refers to the unique material ingredients and process sequence through which the sample was generated. The process recipe contains foreign keys (i.e., references to material nodes) that link information about the device substrate and the solution (the latter contains polymer and solvent information), and to process stage nodes that subdivide the process history (vide infra). Sample and process recipe have a many-to-one (N-1) relationship, as a given sample device can only be associated with one process recipe but the same process recipe could be used for multiple samples. |
64d660394a3f7d0c0d01f530 | 8 | Comparing device data reported from multiple sources requires that the various nuances of the experimental design space are accurately represented in a robust data format. Particularly, understanding the sensitivity of the process space is non-negotiable for the sake of reproducibility and accurate data representation. However, in contrast to the other entities in Figure , it is not straightforward to manifest a data structure that broadly represents the process history for the conjugated polymer layer. This is not only because the real-world process history is extremely complex, but also because the various events in a process history have an explicit order, and the events may not occur consistently across samples. For example, as discussed above, the solution processing procedure may include sequenced pre-treatment techniques to induce polymer aggregation in the solution state. This procedure may include multiple steps, and the ordering of those steps may affect the final film and properties (i.e., sonication then aging, or aging then sonication). The example above exemplifies a broader challenge in robustly handling information in both dynamic and nuanced ways in sample recipes. |
64d660394a3f7d0c0d01f530 | 9 | One avenue to formulate a data structure is to sub-divide the sample generation process into a series of sub-processes that appear in a consistent ordering for any given polymer active layer in an OFET, model the relationships among a standard domain of entities within those sub- processes, and use the resulting graph to sort data. Recently, Walsh et al. proposed a generalized polymer data structure in CRIPT, introducing a data format for process entities that can be sequentially arranged to represent successive material transformations. However, there is no established standard for defining the boundaries of a "process" for the sake of knowledge representation in materials data structures. We propose that incorporating a universal standard to help compose and arrange individual process stages would foster the adoption of generalized data models that can be used to model a broad set of application domains that are sensitive to complex processing histories. |
64d660394a3f7d0c0d01f530 | 10 | Therefore, herein we adopt an international automation standard in ISA-88 to facilitate the conceptual modeling of the conjugated polymer deposition in a logical way (Figure ). 31 ISA-88 is a standard that is used in batch process control to organize the various pieces of data associated with a complex network of instrumentation and process stages, wherein a batch process input material is fed to a defined order of processing actions (e.g., pieces of equipment) in series or in parallel to obtain some output material. Section 4.1 of ISA-88 defines a series of hierarchical subdivisions that are increasingly descriptive of an overall batch process. If the first level of the hierarchy is the overall process, the process is subdivided at the second level as an ordered set of process stages which operate independently from each other, usually in a planned sequence of physical changes in the material being processed. Process stages can be broken up into individual process operations, which are defined as major activities that result in a chemical or physical change in the material inputs. At the lowest level of the process model, process actions represent the minor activities that make up a process operation. Within each level of complexity, entities are a directed set of process sub-nodes organized in serial, parallel, or both. 31 |
64d660394a3f7d0c0d01f530 | 11 | The conjugated polymer process recipe is expanded to process stages of substrate pretreatment, solution processing, film deposition, and post processing at the second level of ISA-88 (Figure ). These entities are fixed stages and represent logical subdivisions that broadly describe the active layer transformation of the input materials. Figure represents a third level of process description, showing that the sequence of process operations that falls within the boundaries of each polymer process stage may vary from sample to sample. Substrate pretreatment refers to all the processing activities related to preparing the patterned device substrate for the deposition process. This sequence may or may not include cleaning treatments (e.g., chemical washing, UV-ozone, plasma treatment, etc.) proceeded by one or more surface modification steps, such as the use of a self-assembled monolayer. Solution processing includes the various ordered steps (e.g., mixing, sonication, cooling, aging, etc.) that transform the polymer/solvent components into the final solution or ink formulation that ultimately gets coated onto the treated substrate during the film deposition stage. The film deposition stage includes only one node that contains information about the parameters and metadata of the coating method (e.g., blade coating, spin coating, drop casting, inkjet printing, etc.). Finally, post-deposition operations such as further chemical treatment or annealing appear under the post process stage node. |
64d660394a3f7d0c0d01f530 | 12 | Further expanding the process operations into individual process actions is possible but provides a level of complexity that may not be required for sufficient database design, as the main purpose of the process model is to identify the most appropriate object classification to store attributes. For example, a "sonication" node might be expanded to process actions such as "set sonication time" or "set sonication intensity", but this information can be just as easily represented by storing the attributes "sonication time" and "sonication intensity" in the parent process operation node (Figure ). However, it should be noted that this next level of process expansion may be useful for relating data commands or readings to computer-integrated or fully automated instrumentation. In either case, all the raw parameters, data, and metadata information are stored in the nodes at the lowest process level. |
64d660394a3f7d0c0d01f530 | 13 | The prior section proposes a general data structure and ontology for storing information related to the formation of the conjugated polymer active layer in a device such as an OFET. ISA-88's sequenced process model also allows for process history to be effectively captured in a data model, as the conjugated polymer is transformed into the final active layer of the measured device through a batch process. The process representation provides a high-level fixed structure (the main process stages of solution processing, substrate treatment, etc.) to promote aspects of a consistent relational schema, while providing flexibility for storing dynamic information within each of the stages. Using the data model described above, an experimental repository of OFET device measurements was curated from a set of published, peer-reviewed literature data (Supporting Information) and unpublished laboratory data. The following section discusses the initial construction and continued curation of experimental device records into the database and provides a brief demonstration of utilizing the database for meaningful searching and data visualization. |
64d660394a3f7d0c0d01f530 | 14 | The database was constructed using PostgreSQL, an open-source, relational database management system primarily based on the structured query language (SQL) that has strong support for NoSQL features. This allowed the database to have the preferred functionalities of the relational model (e.g., data normalization, data redundancy, error checking, etc.), while allowing for storage flexibility where attributes may be dynamic. Preserving relationships is also a key factor in representing sample provenance in a robust manner, which makes certain aspects of the relational model attractive for the sake of interoperability with other community databases. At the same time, PostgreSQL can handle storage and queries on a variety of complex data types, including JSON, XML, and binary objects, which is not a feature that is always available for SQL databases. The mix of SQL and NoSQL features allows the implementation of a data model that provides more convenient and robust organization for structured aspects (process stages, e.g., film deposition) and flexible storage for unstructured information (process operations, e.g., coating parameters, and associated metadata fields). A complete description of the DBMS table schema, based on the data model described earlier, is available in the Supporting Information. |
64d660394a3f7d0c0d01f530 | 15 | Naming errors can in part be mitigated through built-in DBMS functionalities but may persist, especially in JSON formats, due to flexible key-value naming. At the time of writing, the implementation of OFET-db uses some keywords borrowed from other large-scale materials database efforts (e.g., where applicable and available, citation/source keywords from MaterialsMine, process and material keywords from CRIPT, 53 etc.), but largely uses custom keywords that provide more specificity to descriptors relevant to the organic device fabrication domain (e.g., blade, spin, inkjet, dip, etc. to specify different classes of solution coating/deposition techniques). A full list of terms is available in the Supporting Information for the database implementation version described herein. Future design efforts will employ updated terminologies from shared community resources when available, as shared vocabularies promote consistent descriptions and interoperability. An experimental data entry template that incorporates this controlled vocabulary has also been adapted from a similar template shared by MaterialsMine. This template is intended to not only reduce the time and inconvenience that is inevitable for an experimentalist or domain expert to fill out a data record for database entry but provides a tool to reduce the error checking and validation workload on the back end. Future template versions could be implemented as user-friendly webforms, a web application, or directly coupled to electronic laboratory notebooks or integrated laboratory instrumentation to facilitate the process of database inserts for newly curated experimental records. |
64d660394a3f7d0c0d01f530 | 16 | In the future, it is envisioned that a larger population of data can facilitate data-driven knowledge discovery activities through the utilization of data science or machine learning techniques. Figure shows the distribution of charge-carrier mobilities generated using data queried from OFET-db, showing the total spread and statistics of performance values for three relatively well-represented polymers in OFET-db: DPP-DTT, N2200, and P3HT. OFET samples fabricated from all individual polymer types show performance variations that span several orders of magnitude. The higher average and maximum charge-carrier mobility achieved for the DPP-DTT and N2200 data also reflects the general performance advantage of donor-acceptor copolymers versus the model homopolymer P3HT, even despite this large variation. Polymer material characteristics, such as molecular weight and polydispersity, are important factors in performance differences, as demonstrated by Figure . Molecular weight is a well-studied parameter for conjugated polymers in OFETs, and it is generally understood that longer conjugated backbones promote entanglements and aggregates in solution, and thereby enhance long-range molecular order and charge transport pathways in the thin film. The positive correlation between molecular weight and mobility is generally visible for P3HT and DPP-DTT, and variation in mobility for constant molecular weight is visible for all three polymers as similar polymer batches were used across experimental studies. A similar positive trend in mobility is visible for the polydispersity index (PDI), where devices made from a higher PDI polymer are more likely to have mobilities in the upper range of the dataset. |
64d660394a3f7d0c0d01f530 | 17 | Compared to molecular weight, however, the effect of PDI on the charge-carrier mobility is not as well-understood by the community. While this behavior may at first glance be due to a correlation between molecular weight and PDI, a Pearson Correlation analysis shows that these two variables are positively correlated only for P3HT (r = 0.69) but remain uncorrelated for DPP-DTT (0.07) and N2200 (-0.04) (Table ). PDI may be less frequently studied as a tunable experimental parameter since batch characteristics are sensitive to the synthetic procedure, hindering the design of controlled experiments for OFET performance comparisons. Recently, McBride et al. blended different Mw batches of P3HT and found that a wider molecular weight distribution exhibited beneficial effects due to a synergistic behavior between shorter tie chains connecting aggregated domains of larger chains in aged solutions. However, the data analysis above shows that a relationship between dispersity and mobility may be a common trend for copolymer systems, which potentially motivates a broader study into the structural mechanisms behind the mobility dependence on molecular weight distributions. |
64d660394a3f7d0c0d01f530 | 18 | measurements is much smaller than the number of device measurements, but here we show that structural information is available and representative. We expect that future data-driven studies that use OFET-db as a resource will benefit from representative storage of process history, as materials characteristics alone often provide insufficient information for fully understanding the experimental sensitivity of device performance. To that end, with our proposed process ontology we aim to enable the curation of experimental data with the associated process history of samples, with the future intention of providing advanced analyses based on process recipes. A preliminary demonstration of the ability necessary for a broader adoption of representative materials databases, which the work herein addresses for conjugated polymer processing in OFETs. |
64d660394a3f7d0c0d01f530 | 19 | Herein, we demonstrate the design and implementation of a data model for the experimental domain of OFETs as a foundation for broadly enabling database curation and management for organic thin-film electronics. Specifically, capturing process history in a manner that conforms to standard data protocols was a key challenge that was addressed by employing ISA-88, a batch process data model. Then, a database was constructed based on the model using PostgreSQL, enabling storage capabilities for both SQL (structured data) and NoSQL (documentbased data) to provide flexibility without sacrificing the advantages of data redundancy/normalization in the relational model. While the work and discussion presented provides an experimental database that applies to the OFET active layer, it also serves as a model for adaptation to other aspects of organic device experiments by designing the data structure around an accepted process standard. Moving forward, enhancing materials ontologies to comprehensively capture classifications of process steps would facilitate the future design of data models in other domains that accurately manifest the real-world experimental processes. This is a necessary pursuit in elucidating the format in which a sample's provenance is recorded within a database in FAIR data structures. Additionally, future work will build upon the preliminary body of curated experimental data to mobilize data-driven experimentation for polymer-based organic electronics. |
629f5bc52e62693e7f7a8e65 | 0 | Identification of the bioactive structures of endogenous ligands that bind to receptors is important for the development of new drugs and other bioactive compounds. Crystal structure analysis of ligand-receptor complexes is widely used, but is inherently difficult for highly flexible ligands due to the presence of dynamic binding and unbinding equilibria. For example, many of the endogenous ligands of G-protein-coupled receptors (GPCRs) are structurally flexible lipid-based molecules, making it difficult to identify the bioactive binding conformations. Although changes in entropy terms such as solvation and the internal free energy of the interacting species contribute to ligand-protein binding, the direct interaction between the ligand and the protein is the most important factor for specific ligand binding. Three recognition models have been proposed so far: (a) the classical lock-and-key model, (b) induced fit, and (c) conformational selection. In the lock-and-key model (a), the unbound and bound states of the ligand and protein are the same, and there is no change in internal energy upon binding. In the other two models, (b) and (c), binding involves a conformational change in the ligand, the protein, or both. The idea underlying the conformational selection model (c) is that the free energy cost associated with conformational change is small for flexible molecules, and therefore the biologically active conformation is expected to be present in the "unbound" conformational ensemble. |
629f5bc52e62693e7f7a8e65 | 1 | However, there has been opposition to the induced conformational fit mechanism based on the idea that the "biologically active" conformation may be too unfavorable in the unbound state and may only appear as a result of ligand-protein interactions. In general, the conformation of a bound ligand should be stabilized as compared with the unbound ligand; in other words, the interaction energy between the ligand and the binding site (e.g., stabilization by the formation of hydrogen bonds) should be greater than the energy associated with the transition from the unbound to the bound state. Bioactive small molecules (e.g., drugs) usually have a limited number of rotatable bonds, suggesting that a marked dissimilarity between the non-binding and bioactive binding states would be highly unfavorable. However, some highly active endogenous ligands, including lipidic ligands, have a large number of rotatable bonds, and in these cases the occurrence of a bioactive conformation in the conformational ensemble of unbound states seems energetically plausible. In other words, the "bioactive" conformation will always be present in the unbound state, even if the fraction of this conformation is small. Thus, although some authors have argued that adopting a bioactive conformation represents a large energy penalty for a drug, others support the idea that the bioactive conformation is an intrinsic component of the unbound ligand, so that binding involves little energy penalty. The design of efficient ligands always requires a compromise between rigidity, which lowers the entropic cost of binding, and flexibility, which increases the likelihood of favorable interactions between ligand and protein. In this context, conformational restrictions such as cyclization have frequently been used as a strategy to modify the bioactivity of ligands. We considered that this approach could be used to identify a ligand's bioactive conformational space by comparing the bioactivity and accessible conformational space of a series of structurally restricted analogs. To test this idea, we focused here on the endogenous lipidic ligand lysophosphatidylserine (LysoPS) (Figure ). Lysophosphatidylserine (LysoPS) is generated by the enzymatic deacylation of phosphatidylserine (PS). It is an endogenous ligand of GPCRs, and specifically recognizes the three human GPCRs, LPS1 (GPR34), LPS2 (P2Y10) and LPS3 (GPR174). These receptors have roles in various biological functions, including activation/inactivation of the immune system in pathological settings. This panagonistic activity of LysoPS raises the intriguing question of whether the bioactive conformations of LysoPS binding to these receptors are different. LysoPS can be regarded as consisting of three distinct modules (fatty acid, glycerol, and L-serine) connected by phosphodiester and ester linkages. We previously optimized the structure of each module and the ester linkage of the endogenous ligand (1-oleoyl-LysoPS, 1, LysoPS(18:1)), and obtained potent and receptor subtype-selective modulators of GPR34, P2Y10 and GPR174. (MD) also supported the putative GPR34-LysoPS(18:1) docking pose. In addition, we showed that conformational constraint of the glycerol part of LysoPS structure altered the activity toward GPR34 (Figure and Table ). Our present study further confirms that the agonistic activity of LysoPS can be modulated by structural modification of the glycerol part (Table ). These experimental facts are consistent with the idea that LysoPS has a different bioactive binding pose for each of these receptors. |
629f5bc52e62693e7f7a8e65 | 2 | In this context, the aim of this work was to predict the bioactive conformation of endogenous oleoyl-substituted LysoPS(18:1) (1, Figure ) toward the receptor GPR34 by comparing the bioactivities and accessible conformational spaces of four LysoPS analogues (Figure ) conformationally restricted at the glycerol moiety. Previously, we have predicted the selectivity of inhibitor complexes with Bcl-2 and Bcl-XL and the mechanism the resistance of influenza cap-dependent endonuclease against Baloxavir marboxil using molecular dynamics (MD) simulations. Comparing to these simulations, the sampling of accessible conformational space is more difficult because it requires to sample entire conformational space of the target ligand. Hence, this study carried out long-duration replica exchange molecular dynamics (REMD) as well as metadynamics (MTD) simulations to sample comprehensive conformational space of LysoPS and its analogues. The analogues studied in this work were selected from those synthesized during our previous experimental structure-activity relationship study, which established that agonistic potency toward GPR34 was dependent on the regio-and stereochemistry of the glycerol moiety (Figure and Table ). We focused here on the effect of key dihedral angles of the glycerol moiety on the agonistic activity (Figure ). |
629f5bc52e62693e7f7a8e65 | 3 | The results were compared with the reported dihedral angles of LysoPS(18:1) docked to a homology model of GPR34. Although the crystal structure of GPR34 has not been determined, this docking structure was validated based on the SAR of LysoPS derivatives. Figure . Definition of dihedral angles of the glycerol moiety targeted using REMD and |
629f5bc52e62693e7f7a8e65 | 4 | In our previous study, a cyclic scaffold, 2-hydroxymethyl-3-hydroxytetrahydropyran, was introduced instead of the glycerol moiety to reduce the flexibility of LysoPS and simultaneously to modulate the activity towards GPR34 (Figure ). The carbon atoms of cyclic LysoPS were numbered to correspond to the glycerol framework of LysoPS, from carbon 1, linked to the acyl chain (18:1) to carbon 3, connected to the phosphoserine part (we call this a 1a-2p type analogue: 1a represents the acyl chain at the primary (1°) alcohol and 2p represents the phosphoserine part at the secondary (2°) alcohol (Figure ). |
629f5bc52e62693e7f7a8e65 | 5 | If the primary alcohol is bonded to the phosphoserine part and the secondary alcohol is linked to the acyl group, we refer to the compound as 1p-2a type. Thus, the glycerol moiety was fixed by connecting the hydroxyl group of 2-glycerol alcohol to carbon 1 or 3 to provide four different core scaffolds of cyclic LysoPS analogues, including cis and trans isomers (Figure and Figure ). The fatty acid moiety is oleic acid (18:1), as in endogenous LysoPS (18:1) (1). The agonistic activities of the cyclic LysoPS derivatives were evaluated by means of TGF shedding assay. The results showed that both cis and trans core scaffolds having the acyl chain at the primary alcohol position of the tetrahydropyran ring were potent and selective for GPR34 (Table ). For the present computational analysis, we selected trans-1a-2p (EC50 = 100 nM, GPR34) and cis-1a-2p (EC50 = 210 nM, GPR34) as active ligands, and trans-1p-2a and cis-1p-2a as non-active ligands, together with endogenous LysoPS(18:1). |
629f5bc52e62693e7f7a8e65 | 6 | Forty replicas were constructed over the temperature range from 300K to 500K at intervals of 5K in order to equilibrate the system for REMD and each set of replicas was calculated for 8 µsec (8000 nsec). The results showed that LysoPS(18:1) can occupy a broad conformational space with an extensive distribution of dihedral angles around ±60° and ±180° in both the D1 and D2 axes, whereas the distributions of the cyclic LysoPS analogues were spatially restricted (Figure and Figure ). |
629f5bc52e62693e7f7a8e65 | 7 | showed conformational distributions with dihedral angle 2 around +60° and +180°, but not with dihedral angle 2 in the range from 0° to -180°, because of the rotational restriction caused by changing glycerol to a tetrahydropyran ring (Figure ). In the case of the cis isomers, the histogram of cis-1a-2p displayed conformations with dihedral angle 1 around ±60°, but there are no low-energy conformations with dihedral angle 1 around ±120° to ±180° (Figure ). On the other hand, cis-1p-2a showed conformations with dihedral angle 2 around ±60°, while there were no conformations in the ranges from +120° to +180° and from -120° to -180° (Figure ). These REMD results are summarized in Table , where yellow highlighting indicates accessible conformations and a mark indicates possible bioactive ligand conformations. In the case of the bioactive compounds trans-1a-2p and cis-1a-2p, the accessible conformations will include the bioactive ligand conformation, whereas in the case of inactive trans-1p-2a and cis-1p-2a, the bioactive conformation would not overlap with the accessible regions. Table indicates the number of overlaps of the four cyclic LysoPS analogs with respect to the two dihedral angles (yellow, one; orange, two; pink, three; red, four), where "overlap" is defined to occur where the active analogs (trans-1a-2p and cis-1a-2p) sampled possible bioactive conformations (marked by '✔'), and where the inactive analogs (trans-1p-2a and cis-1p-2a) did not sample the bioactive conformations. The red boxes (+60 and -180) contain the active compounds, but do not contain the inactive compounds. Thus, these REMD results suggested that the dihedral angle range of around +60° (D1) and -180° (D2) may contain the bioactive ligand conformation for GPR34. |
629f5bc52e62693e7f7a8e65 | 8 | Metadynamics simulations (MTD) were performed and the free energy landscapes of LysoPS(18:1) and its derivatives (trans/cis-1a-2p and trans/cis-1p-2a) were constructed with respect to the dihedral angles CV1 and CV2 (CV1: green, corresponding to the dihedral angle 1 (D1) for REMD; CV2: orange, corresponding to the dihedral angle 2 (D2) for REMD) (Figure ). Energy landscapes were obtained at 300K (Figure ), and red to blue colors represent low to high free energy states. For LysoPS(18:1), the minima MTD was also carried out for the four cyclic LysoPS derivatives and the corresponding landscapes were obtained (Figure ). The minima of trans-1a-2p are located at around 60° and 180° along the CV1 axis (lower than 5.0 kcal/mol, red region), while the free energy of the regions along CV1 from -180° to 0° is extremely high (20.0 kcal/mol) because cyclized glycerol blocks single bond rotation (white region, Figure ). Similarly, the minima (lower than 5.0 kcal/mol) for cis-1a-2p (red regions) along CV1 were located around ±60°, and very high free energy states (white color region) were located around ±180° to ±120° (20.0 kcal/mol) along CV1 (Figure ). Both trans-1a-2p and cis-1a-2p are bioactive toward GPR34. In the case of inactive trans-1p-2a, the energy minima (lower than 5.0 kcal/mol) along CV2 were located at around 60° and 150°, and the range between 0° to -180° (CV1) was a high energy region (20.0 kcal/mol) (Figure ). The MTD of cis-1p-2a, which is also inactive toward GPR34 with EC50 >1 µM, showed energy minima along CV2, located around ±60°, while the area from ±180° to ±120° along CV2 is high energy (Figure ). In the case of inactive trans-1p-2a and cis-1p-2a, the bioactive conformations should lie outside the occupied regions, that is, they should lie in the high energy regions. These results confirm that the tetrahydropyran structure restricts the accessible conformations of the molecule due to prohibition of ring flipping and singlebond rotation. While the populations of minor conformers are different between the REMD and MTD calculations, the emergence of conformers is consistent in the two calculations. From the MTD results, the region around CV1 = +60° and CV2 = -180° is consistent with conformations of bioactive cyclic LysoPS derivatives. This region is in agreement with the range of dihedral angles (red box in Table , D1 = +60° and D2 = -180°) obtained by REMD (yellow dashed circle in Figure ). In other words, the conformation involving the coordinates +60° (D1, CV1) and -180° (D2, CV2)) (black dashed circle in Figure ) should be responsible for the bioactivity of LysoPS(18:1) and its cyclic analogues towards GPR34. |
629f5bc52e62693e7f7a8e65 | 9 | In the most stable binding pose of LysoPS(18:1) to GPR34 (Figure ), the dihedral angles corresponding to D1, D2 and CV1, CV2 of the glycerol moiety are +99.8˚(corresponding to D1 and CV1) and -155.2˚(corresponding to D2 and CV2) (Figure ). These dihedral angles are consistent with both the distributions provided by REMD (Figure ; yellow dashed circle) and the stable energy states obtained by MTD (Figure ; black dashed circle). However, it should be noted that short-duration REMD did not detect the minor conformer of LysoPS(18:1) that finally turned out to be the bioactive conformer. Nevertheless, the results suggested that the bioactive conformation of LysoPS(18:1) bound to GPR34 is present among the unbound ligand conformations. Intriguingly, this bioactive conformation is located in a minor distribution region, but would be selected upon binding to GPR34. The energy penalty to take a minor conformation, that is, the energy difference between the energy minima conformation (G=0.00 kcal/mol in MTD, e. g. CV1=177.46°, CV2=-172.87°) and the conformation in MTD approximately corresponding to the binding pose of LysoPS(18:1) (CV1=99.66°, CV2=-155.59°) is 4.26 kcal/mol, which will be compensated by binding to GPR34. |
629f5bc52e62693e7f7a8e65 | 10 | The conformational populations with respect to two dihedral angles, D1 and D2, obtained by REMD (Table B) and the energy landscapes (Figure ) constructed with CV1 and CV2 afforded by MTD gave consistent distributions of conformations with respect to the glycerol moiety for all four conformationally constrained analogues as well as LysoPS(18:1) (Figure ). Comparison of the conformational distributions of the active and inactive analogues allowed us to identify the positioning of the bioactive ligand conformation (coordinates at around +60˚ (D1, CV1), -180˚ (D2, CV2)). This bioactive conformational region is among the less populated of the accessible regions of LysoPS(18:1) itself (Figure ). Notably, the measured dihedral angles of LysoPS(18:1) |
629f5bc52e62693e7f7a8e65 | 11 | in the docking pose with GPR34 that we previously reported are included in the range of dihedral angles of bioactive ligand conformations obtained from REMD and MTD calculations. In other words, the two dihedral angles of the glycerol moiety in the bound conformation of LysoPS(18:1) to GPR34 correspond well to accessible conformations of the unbound ligand. |
629f5bc52e62693e7f7a8e65 | 12 | Sketcher in Maestro (version 2017-4) (Schrodinger Inc., U.S.A.) and each structure was ionized completely, that is COO -, NH3 + and P-O -. The 3D structures were built and saved as 3D project entries by using 2D Sketcher. These 3D structures were submitted to System Builder Panel to be charge-neutralized by adding a single Na + atom. Then, each molecule was embedded in a 10Å x 10Å x 10Å water (TIP3P) solvent box and the system was buffered with 0.15 M NaCl with using the OPLS3 force field. The whole system (molecule + Na + + water + NaCl) was relaxed by a simple molecular dynamics calculation with Desmond (version 2017-4) (Schrodinger Inc., U.S.A.) for 5 nsec using the OPLS3 force field at 300K. The structure after 5 nsec relaxation was used as the initial structure for the following replica exchange calculations and metadynamics calculations. |
652668e78bab5d20551ad1b7 | 0 | We present a battery screening Python pipeline, VOLTA. It allows for a novel battery active material explorative workflow, prioritizing the cell level performance indicators, such as cell capacity and voltage profile. This is achieved by the construction of a starting dataset of both observed and virtual active materials from the Materials Project, the implementation of the physics-based ARTISTIC project pipeline for the assessment of practical electrode properties like porosity and thickness, and the coupling of the electrodes into virtual cells, whose figures of merit are calculated, like voltage and capacity. The screening can be conducted by applying filters to these cell-level properties, achieving a indirect selection of the most suitable active materials. |
652668e78bab5d20551ad1b7 | 1 | The approach is validated through comparison to current commercial battery technology, and we demonstrate that VOLTA is able to identify promising electrode materials for high energy batteries, like the wellknown LiCoO 2 , LiNiO 2 and graphite. We also illustrate a case-study, where the pipeline is used to identify suitable low-voltage, realistic virtual batteries obtained by combining entries of the Materials Project battery database (a battery revealing type of task). |
652668e78bab5d20551ad1b7 | 2 | Material discovery has traditionally been carried out via a scientifically informed trial and error. Such an approach relies on the formulation of successive hypotheses oriented towards the restriction of the exploration space on the basis of the previous observations. Such a strategy relies heavily on experimentation and on domain knowledge, and while this can be beneficial for local optimization, it can also result in biased research. Nonetheless, materials science has relied on this approach for centuries. A new strategy in material discovery was introduced in the last decade with the use of descriptors-driven, high throughput screening. Any descriptive parameter (or set of) can fall into this category, but one might divide them into anthropic (i.e. human-readable) and data-driven descriptors, the latter being the most recent to be introduced. |
652668e78bab5d20551ad1b7 | 3 | Anthropic descriptors have been developed to rationalize and predict materials properties from domain knowledge. Examples include proxies for stability, e.g. the energy above the convex hull, or more specific descriptors like the Li phonon-band center for Li-ion conductor materials. [3] While such a modeling strategy has proven useful, finding correlation between a certain descriptor and a property does not necessarily imply causation: screening with anthropic descriptors always needs some caution, as pointed out by Ghiringhelli et al. This classic screening strategy mainly relies on Density Functional Theory (DFT), thanks to its high accuracy and its capability to output readable, physically meaningful descriptors, like formation energies. The high computational cost of DFT does not allow for the direct exploration of a wide chemical space, hence funnel-like screening pipelines had to be developed to progressively reduce it. Material screening for energy storage applications, and batteries in particular, can naturally be conducted with such an approach. Traditional computation-based science is giving a tremendous impulse to an accurate and high-throughput prediction of battery-relevant materials' properties. |
652668e78bab5d20551ad1b7 | 4 | As an example, one could consider the sheer amount of information (i.e. anthropic descriptors) that is hosted on e.g. the Materials Project (MP) for a number of electrode materials: specific and volumetric capacities, volume changes upon (de)lithiation, the whole voltage profile of each electrode material, together with other relevant, less domainspecific properties like stability. All this is calculated by first principle methods. Databases like MP or NOMAD provide a significant step forward in terms of efficiency and provide homogeneous data-sets both opening to statistical analyses on materials' anthropic descriptors, and to a new screening paradigm: the use of data-driven descriptors. |
652668e78bab5d20551ad1b7 | 5 | Data-driven descriptors are generally achieved and employed by the means of machine learning (ML), and they can be useful when trying to reduce the computational cost of modeling tasks. Moreover, data-driven descriptors might both avoid human biases and allow for the analysis of multi dimensional relations between descriptors and real properties. Examples of these descriptors are the weights and biases of neural networks, whose iterative optimization (i.e. training) can lead to predictions comparable in accuracy to the ones of DFT. Such accuracy, though, comes at the cost of model interpretability, albeit some strategies can be used to improve on that like the use of increasingly "expressive" material encodings. The use of ML is not limited to property prediction, but it branches out to the use of text mining to retrieve information about synthesis methods and conditions and the use of neural networks to predict the synthesis path of target compounds, illustrating the high flexibility of this kind of approach. As pointed out by Ling et al. the emergence of ML has brought a sensible push forward in general purpose material science: from reaction and synthesizability network analysis [15] to the acceleration of "traditional" force field calculations and material discovery as a whole. The field of property (i.e. anthropic descriptors) prediction can benefit from ML models as well, in taking advantage of the cheaper computational costs of the models training as compared to their calculation, that is essentially based on modeling with data-driven descriptors. [9] Battery modeling, though, can hardly rely only on first principle cal-culations to achieve a "holistic" description of a whole galvanic cell. |
652668e78bab5d20551ad1b7 | 6 | In fact, the computational costs for calculating observables like energy barriers for solid-state diffusion or diffusivity within the whole electrode can get prohibitive, given the vast configurational space, thus making the complete modeling of all the present processes very costly. Increasing the scale of modeling, two main domains need to be addressed: the mesoscale and the macroscale. At the mesoscale level, ML has already proven to be up to the task when predicting the outcomes of electrode formulation and manufacturing process, from the slurry phase to the final calendering. [24] Such predictions rely heavily on high-quality training data, either experimental or output from physics-based models, namely the ARTIS-TIC pipeline , which was used in the present work to predict electrode thicknesses and porosities. A similar reliance on training data was highlighted for mesoscopic modeling of electrode dynamics upon battery cycling . |
652668e78bab5d20551ad1b7 | 7 | On the other hand, multiscale modeling proves useful to analyze the macro, millimeter scale, too. This is quite intuitive insofar as we assume that the macroscale is fully informed by the physico-chemical processes lying beneath it, although taking into account the emergence of new properties from the combination of the simpler ones. While a fascinating approach, it seems incompatible with the screening of active materials (AMs) for electrodes: the chemical, structural and textural properties of AMs can indeed vary greatly and quickly become unmanageable in a multi-scale model. |
652668e78bab5d20551ad1b7 | 8 | Inspired by these considerations, we developed a first-of-its-kind (to the best of our knowledge) algorithm for battery AM screening by working on the resulting cell-level properties when the compounds are coupled in practical galvanic cells. The software is called VOLTA. Hereafter, we will describe its workflow, discuss its validity and apply it to a battery revealing case study. |
652668e78bab5d20551ad1b7 | 9 | Figure : pictorial representation of the VOLTA workflow. The color coding in the background specifies that all the decisions and calculations on the left (orange) side are relevant to the electrode simulation part of VOLTA, conducted with the ARTISTIC workflow . ARTISTIC is the external pipeline used to simulate the electrodes manufacturing and obtain realistic mass loading and porosity values. On the right side, the color coding shows how the pipeline starts from AM-level (light blue) decisions (filtering), operations (AM coupling) and calculations (the descriptors reported in red). These descriptors are further illustrated: for the feasibility level, see Table , for the voltaic ones, see Figure . Finally, the pipeline achieves the cell level (green), allowing for a new filtering decision. |
652668e78bab5d20551ad1b7 | 10 | Figure displays the general workflow of the algorithm. The first step consists of the collection of the so-called "insertion electrode objects" from MP, i.e. Python objects containing a framework material (typically in its oxidized state) and phases obtained by virtually intercalating ions like Li + , Na + or other ions into the framework (reduced states). We will refer to these Python objects containing active material data as AMOs (active material objects). For each of the AMOs, numerous batteryrelevant properties have been calculated by MP, such as their voltage profiles upon (de)intercalation, their capacities and the stability of the phases, and are organized by VOLTA into a Python pandas data-frame, for convenience. A complete list of the available AMO-level descriptors is available in SI, as well as on the MP website itself. |
652668e78bab5d20551ad1b7 | 11 | The AMOs pass through a filter, first, to allow the user to exclude a priori materials they might deem not worth considering. For instance, one might need to exclude all the AMOs above a certain energy above the hull threshold, or those operating outside the electrochemical stability window of a given electrolyte. It is worth mentioning that not all the AMOs gathered from the MP have been observed and reported in literature: these phases can appear as virtual intermediate phases of known materials or as fully virtual AMOs, thus making the screening pool potentially very diverse. It is also an option to exclude all the virtual phases, ending up with a more limited pool of only observed materials. The following step combines each AMO with all the others (without repetition), creating a pool of virtual cells composed of a positive and a negative AMO. The positive side being the AM operating with higher average voltage. For each of these virtual cells, VOLTA calculates multiple cell-level descriptors, that can be classified into three categories: the cell feasibility descriptors, the voltaic descriptors and the capacityrelated ones. A complete list of all the descriptors is available in SI, and the main ones will be discussed hereafter. Lastly, the software allows the user to filter the virtual cells and to score them by any of their characteristic, identifying e.g. the ones yielding higher energy density. |
652668e78bab5d20551ad1b7 | 12 | A detailed exposition of the cell-level descriptors, assuming the positive electrode is assigned to the fictitious AM with composition Li 0-1 C and the negative one to Li 0-1 A, follows hereafter. For the sake of this example, let us assume that Li 0-1 C is only defined by its lithiated and delithiated phases (C, LiC), and that Li 0-1 A comprises the A, Li 0.5 A, LiA phases. We will refer to the fully lithiated and fully delithiated phases as end-member phases. Table : illustratory phase-level metrics. |
652668e78bab5d20551ad1b7 | 13 | The three feasibility metrics enable the user to assess how viable a virtual cell would be to put together in practice, and are based on the end-member phases' energies above the hull and whether they have been observed experimentally. Example values are reported in Table , and their respective definitions are: |
652668e78bab5d20551ad1b7 | 14 | 2. the cell possibility: this boolean metric evaluates to true if the virtual cell contains at least one observed end-member phase for each side. The two observed phases also need to be one in its lithiated and the other in its delithiated state. In the example above, this boolean does evaluate to true; |
652668e78bab5d20551ad1b7 | 15 | 3. the cell stability metric (SM): this metric is given by the sum of the SD calculated for the positive and negative side of the cell. Partial SD equals zero if at least one of the phases has been experimentally observed; if not, it equals the lowest energy above the hull (EAH) value among those of the end-member phases. In the exemplary cell reported in table 1, SD would amount to 0.02 eV. |
652668e78bab5d20551ad1b7 | 16 | • the voltage spread: the difference between the voltage extremes. To estimate the cell capacity, VOLTA calculates the capacity of both the positive and negative electrodes; the lower of the two limiting the capacity of the cell overall. While it is easy to prove that cell capacity is directly proportional to the product of the AM's specific capacity and of its density (eq. 1), the goal of VOLTA is to provide reasonable experimental-like values, to facilitate the comparison between theoretical and real world cells. |
652668e78bab5d20551ad1b7 | 17 | The next step is to assume an electrode formulation, assuming that the electrode could be manufactured by a mix of AM, binder and conductive carbon cast, dried and calendered on the current collector substrate. Here VOLTA distinguishes between electron-conducting and electroninsulating AMs, by applying an AM-rich formulation for the former (96-2-2) and a binder plus conductive carbon for the latter (85-7.5-7.5); these values are, however, user-definable. The distinction between electronic conductors and insulators is based on the DFT calculated average band gap of the phases that appear in each electrode upon cycling. The threshold value can be provided by the user, but with a caveat. The band gaps utilised have been calculated by MP with PBE-DFT , and corrected with the empirical method pointed out by Morales-García et al. . Such a correction is needed to tackle the systematic underestimation of band gaps due to the PBE level of theory. Despite the relatively low amount of training data, the method seems to predict well the experimental band gaps falling in the 1 -5 eV interval. Most of the corrected band gaps of the materials analysed in the present work fall in this interval. |
652668e78bab5d20551ad1b7 | 18 | At this point, the only variables left are the electrode porosity ϵ and the three mass values (active material m AM , conductive carbon m C , binder m B ). The ARTISTIC project Online Calculator was used to calculate porosity and mass of active material as follows: given an electrode formulation, the particle size distribution of the AM and other inputs (listed in SI), the pipeline simulates the whole electrode manufacturing process (slurry mixing, drying, calendering) and returns the electrode's mass loading per area, and its porosity. [32] [34] This simulation step is aimed at providing a more reasonable model of the electrodes' microstructures, hence their mass loading and porosity. |
652668e78bab5d20551ad1b7 | 19 | All these parameters are then substituted into the system of equations, and the required masses of conductive carbon and binder are calculated, together with the electrode thicknesses (hence volumes), and the electrode capacity is calculated. At the cell level, i.e. when two AMOs are coupled and all the metrics above are calculated, the electrode with lower capacity will dictate the capacity of the cell. A complete list of electrodes and cells descriptors is reported in Table |
652668e78bab5d20551ad1b7 | 20 | Cell desirability D is an arbitrary definition of the general attractiveness of a virtual cell. It takes into account three components: the cell's performance (P), user definable as its capacity or energy, the cell's overall stability metric (SM), mentioned above, and the cell's sustainability score (CESI), calculated along the lines of the Chemical Element Sustainability Index (CESI score). Being that this sustainability index is not defined for some chemical elements, it was decided to assign a CESI score to each material based on the transition metals (TMs) it contained. When multiple TMs were present, the average CESI score of the TMs was chosen to represent the material, in order not to penalise apriori the multinary materials. All the descriptors are normalised, so that 0 ≤ D ≤ 3, with D = 3 indicating the most desirable cell possible. The single descriptors are multiplied by "importance factors" (p, c, sm), that are user definable. Choosing, e.g., c = 0 would screen cells ignoring their sustainability score, or doubling sm would double the weight the user attributes to the cells' AMs stability. Lastly, the user can tune the performance metric with its weight p. |
652668e78bab5d20551ad1b7 | 21 | Two approaches have been followed to validate the results achieved by VOLTA. First, the software was ran to identify the virtual cells yielding the highest energy: VOLTA outputs well-known, commercial AMs, henceis capable of identifying some of the best battery materials. Moreover, the frequency of appearance of some commercial AMs was analysed as to seek whether they would appear more often when considering virtual cells with increasing energy density. Li 0-1 CoO 2 (LCO), Li 0-1 NiO 2 (LNO) and Li 0-1 FePO 4 (LFP) (whose respective battery IDs within MP are: mp-22526_Li, mp-25210_Li, mp-19017_Li ) were identified as suitable AMs to keep track of, being well known, commercially available positive electrode active materials (PEAMs). |
652668e78bab5d20551ad1b7 | 22 | As expected, LCO and LNO display high theoretical specific capacities within the electrodes dataset, which contains a total of 2476 electrode materials. LFP's capacity still scores well above the median value, but is distinctly lower than the LCO and LNO's. Coming to their average voltages, all of them are well above 3 V (see Figure ). |
652668e78bab5d20551ad1b7 | 23 | To conduct a similar evaluation on negative electrode active materials (NEAMs), we realised that the battery section of MP did not contain, to date, entries for commercially available ones. We then proceeded to manually add two entries, marked with an asterisk, titled graphite (Li0-1C6*) and LTO (Li4-7Ti5O12*). Their specifics are reported in Table in SI. When compared to the other materials, graphite performed exceptionally well in terms of theoretical specific capacity, scoring in the top 99.4 percentile. Its exceptionality as a NEAM is then further confirmed by the outstandingly low average operational voltage: it indeed scores as the lowest of all in the dataset. LTO, on the other hand, seems to make for a reasonably good NEAM, scoring in the top 77.9 percentile by theoretical specific capacity and in the 4th percentile by operational voltage. It is important to notice, though, that this rather small shift from 0th percentile to 4th is accompanied by an increase in average operational voltage of more than 1.5 V. This is due to the low number of AMs operating at voltages below 1.5 V, as can be observed in Figure (a). The AMs specific capacity distribution is visualized in Figure (below). Numeric values are also reported in Table . Before running the VOLTA pipeline, the AMs were filtered, in order to exclude the phases too unstable and the AM operating outside the practical stability window of a conventional liquid electrolyte, i.e. 0 to 4.5 V The number of AM left was 2028. Moreover, the electrode construction parameters are the same as in Section 4.1: the electrodes' thicknesses are calculated from the mass loadings and porosities output by the Artistic calculator (reported in Table in SI). After the cell building step, only the possible cells were considered (Possibility boolean evaluating to True, see Section 3). The cells with voltage spread above 0.6 V were also excluded, not to exceed the stability window of the electrolyte too much. Table reports all the filtering criteria for AMs and for the constructed cells. |
652668e78bab5d20551ad1b7 | 24 | The resulting electrode thicknesses (calculated as per Section 3) are distributed between 49.5 and 12.8 µm (PEAM) and 48.8 and 11.5 µm (NEAM). Given the distribution of AMs specific capacities visualized in Figure (b), and given the fact that the virtual cells are built by mere combination of AMs, it follows that the resulting cells can vary substantially in terms of capacity balancing, i.e. it is likely that materials with very different specific capacities are coupled in a virtual cell, resulting in turn into electrodes yielding different capacities per unit of area. This results in a capacity balancing distribution between a mini-Range (AMs) Range (cells) Average voltage (V) 0 -4. Half of the capacity balance data points fall within the range 0.49 -1.70. The absence of a cell balancing routine has to be seen as a feature of VOLTA: the electrode manufacturing simulation output by the ARTISTIC pipeline is computationally expensive, and in order to balance a cell one would need to iteratively run such a simulation until converging to the maximum mass loadings allowed by electrode geometry, porosity and the chosen formulation. The single-point electrode simulation was deemed more desirable than a cell balancing approach based on assumed porosity and electrode thickness. |
652668e78bab5d20551ad1b7 | 25 | In order to further validate the approach one could run the algorithm, order the virtual cells by energy, and calculate the share of virtual cells that contain certain AMs, for instance the commercial materials we introduced in Table . This calculation could be repeated for the cells falling within the top n percentile with n assuming the values of (100), 80, 50, 20, 10, 5 and 1. One could expect the frequency by which AMs with high capacity and high (PEAMs) or low (NEAMs) operating voltage appear to increase considering increasingly energy dense cells. The calculated rates of appearance are reported in Figure . |
652668e78bab5d20551ad1b7 | 26 | LCO and LNO display a clear monotonic growth, signifying that for increasing cell energy an increasing share of cells contains either of these as positive electrode active materials. It is also worth mentioning that their volumetric theoretical capacities score in the top 99th percentile. The observed trend is reasonable, being that such electrodes score so well in terms of theoretical capacities and display a rather high average operational voltage. LFP, on the other hand, displays a monotonic decrease, that plummets when considering the top 1 percent of cells. Being that LFP's average operational voltage is close to the dataset mean (45.9th percentile), the trend can be justified by the lower theoretical specific capacity (78th percentile), and theoretical volumetric capacity (74th percentile) typically associated with species like polyanionic compounds. |
652668e78bab5d20551ad1b7 | 27 | Coming to NEAMs, their appearance frequency values are plotted in Figure . A monotonic increase in appearance frequency was observed for both graphite and LTO, confirming that the software captures indeed known, chemical common-sense trends. Moreover, graphite displayed a substantial increase in frequency, almost reaching a frequency of 5% in the first percentile of cells by energy (i.e. 5% of the cells in the first percentile by energy have graphite as negative electrode active material). |
652668e78bab5d20551ad1b7 | 28 | A similar analysis can be conducted by ordering the virtual cells by desirability, as defined in equation 6 in Section 3. The cell's performance (defined here as cell energy), the cell's sustainability score and its stability metric were weighted equally (p = 1, c = 1, sm = 1). The frequency of appearance of the commercial PEAMs is plotted in Figure (c). The overall picture seems different from Figure (a), diplaying a clear reduction in overall frequency of LCO and in the frequency peak at the top percentiles. On the other hand, LFP and LNO seem to grow dramatically in frequency upon considering higher desirability cells. It is difficult to deconvolute the contributions of the three components directly, but the main observation is that LCO is penalised in terms of CESI score. Its normalised CESI score is 0.32, while LNO and LFP score respectively 0.61 and 0.75. The materials have all SD = 0, as at least one of their end-member phases have been observed, and their relative performances have been discussed above. The frequency of LCO is greatly reduced, and the peak relative to the top 1% cells only reaches 0.1%, half of the value reached when considering only performance. This results in a reduction in desirability of Co-containing LCO, despite its good performance, in accordance with the current trend of Co-removal from PEAMs. LNO displays a growth trend right away, that reaches a slightly lower value than in the previous analysis (slightly more than 0.2% vs 0.25%). This can be interpreted as a slight devaluation of LNO when sustainability is introduced in the analysis. LFP, on the other hand, undergoes a severe change in frequency behaviour: from a relatively undersirable compound in the pure-performance analysis, it is promoted to very high frequencies, comparable in behaviour to LNO. These observations are in line with the difference in sustainability between iron and elements such as nickel and cobalt. This demonstrates that VOLTA can also be applied to battery screening problems not only limited to the maximisation of the virtual cells performance, making it a rather flexible algorithm. |
652668e78bab5d20551ad1b7 | 29 | Figure : PEAM (a) and NEAM (b) frequency analysis for virtual cells ordered by energy; the frequency of appearance of well-known, commercial battery AMs follows a reasonable increasing trend upon considering ensembles of cells with increasing energy density; (c) PEAM frequency analysis for virtual cells ordered by desirability, considering equally performance (i.e. energy), sustainability and the cell stability metric. |
652668e78bab5d20551ad1b7 | 30 | After running VOLTA with the constrains reported in Table , i.e. to reveal virtual cells with maximum energy density, limited voltage spread and Possibility evaluating to True (the PEAM is marked as observed in the MP in its fully charged or fully discharged states, and the NEAM is marked as observed in the opposite state of the PEAM). |
652668e78bab5d20551ad1b7 | 31 | At first glance, we observe how relevant graphite is as a NEAM in this battery revealing problem. Its low operational voltage and high theoretical capacity, which we discussed above, make it a good NEAM to match with a number of possible PEAMs. On the PEAM side, we observe mainly transition metal oxides, including the commercial, wellknown Li 0-1 CoO 2 . Two different Li 0-1 MnO 2 entries appear, too. These are referred to two separate MP elecrode objects(mp-18767_Li and mp-33009_Li), thus highlighting that VOLTA does not explicitly take into account crystallographic information. Such an approach is not to be considered as "composition-only", though, as VOLTA distinguishes polymorph AMs; it becomes necessary, though, to determine from the MP database the structural differences. In the present case, mp-33009_Li is obtained from virtual lithiation of rutile-structured MnO 2 (mp-510408, observed), while mp-18767_Li is obtained from the virtual delithiation of orthorombic Li 0-1 MnO 2 (mp-18767, observed). The highest performing cell is obtained by combining graphite and Li 0-2 CrF 4 . This reasonably follows from the intuition that high capacity is also linked to multiple reduction steps on the active transition metal and that voltage can be increased by inductive effect shifting from oxides to halides. The ability of VOLTA to identify likely battery candidates is thus demonstrated. The MP entry IDs of the identified materials are reported in Table in SI. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.