id
stringlengths 24
24
| idx
int64 0
402
| paragraph
stringlengths 106
17.2k
|
|---|---|---|
668d80145101a2ffa8e358c3 | 10 | Self-Consistent Field Spin-Orbit (CASSCF-SO) calculations (Selected parameters compiled in Table , see SI Figures S58-S69 for all magnetic data). In general, there is reasonable agreement between the measured and calculated susceptibility values as well as those for the respective free ion values (Sm(II) 4f F0; Eu(II) 4f S7/2) for all complexes, though calculated values for 1-Sm, 3-Sm and 4-Sm are slightly higher than predicted due to the thermal population of the lowest-lying FJ excited states; the shapes of the calculated curves are accurately reproduced. Complex 1-Sm exhibits a ฯMT value of 3.09 cm 3 mol -1 K at 300 K, which is in close agreement with the free ion value for two Sm(II) ions. |
668d80145101a2ffa8e358c3 | 11 | Complexes be due to a magnetic interaction between the two Eu(II) ions. For 1-Eu and 3-Eu, the experimental values lie approximately 1.6 cm 3 K mol -1 above that predicted by CASSCF, whereas there are closer agreements with the experimental and calculated values for 4-Eu, which differ from those predicted by 0.18 cm 3 K mol -1 . Analogously to 1-Sm, complex 1-Eu exhibits a ฯMT value of 17.35 cm 3 mol - 1 K at 300 K, which is in close agreement with the free ion value for two Eu(II) ions; the low temperature magnetization data are also consistent with spin-only ions. Hence, we focus our discussion on the solid-state spectra. The spectra for the monometallic complexes were simulated, using EasySpin, with the simple Zeeman and ZFS spin Hamiltonian: |
668d80145101a2ffa8e358c3 | 12 | where S is the electron spin quantum number, ๐ ๐ต is the Bohr magneton, B is the applied magnetic field, g is the electronic g-value (treated as isotropic), and D and E are the axial and rhombic components, respectively, of the ZFS interaction matrix. Higher-order ZFS terms, possible for S = 7/2, were neglected. |
668d80145101a2ffa8e358c3 | 13 | Powder samples of 3-Eu at Q-band and low temperature give spectra with intense resonance ranging from ca. 0.5 T (with weaker features at lower field) to the maximum range of the electromagnet (1.7 T). In the wings of the spectrum a regular progression of transitions with separation ca. low-field resonances it was necessary to assume a slight preferential ordering of the system along the molecular z-axis in the magnetic field (as opposed to a true powder average); this is consistent with a negative D (hence easy-axis magnetic anisotropy). The EPR spectra of dimetallic 1-Eu are more complex at low temperatures, and on warming intensity rapidly grows in the middle of the spectrum (see SI Figures ). This is only consistent with an exchange interaction between the two Eu(II) ions. Modelling of low temperature magnetic data shows that this exchange must be very small (|J| < ca. 30 GHz), but this puts it in the same regime as the Eu(II) ZFS, hence very complex spectra result and we have as yet been unable |
668d80145101a2ffa8e358c3 | 14 | The electron densities of 3-Yb and 4-Yb were calculated using PBE to determine the nature of the Yb-P interactions (see SI Figures ). Frontier molecular orbitals show that there is no substantial orbital overlap between Yb and P (Figures ), where the lone pair remains localized on P, although a larger P 3p contribution is present in the highest occupied molecular orbital of 3-Yb. Delocalization indices, ๏ค(Yb-P) and Wiberg bond orders (WBO) report the number of electrons shared between two nuclei; for 4-Yb, the average ๏ค(Yb-P) = 0.24 and WBO is 0.28, however these values are higher for 3-Yb: ๏ค(Yb-P) = 0.33 and WBO is 0.36. The electron density at the bond critical point ๏ฒBCP(Yb-P) shows the Yb-P interaction in 3-Yb (0.034 a.u.) is stronger than in 4-Yb (0.015 a.u.). |
668d80145101a2ffa8e358c3 | 15 | All manipulations were conducted under argon with the strict exclusion of oxygen and water by using Schlenk line and glove box techniques. K{P(SiMe3)2} (KPโฒโฒ) and LnI2(THF)2 (Ln = Sm, Eu, Yb) were prepared following literature procedures. Diethyl ether and toluene were purged with nitrogen and passed through columns containing alumina catalyst and molecular sieves and pyridine was dried over CaH2 before being degassed, refilled with argon and stored over 4 ร
molecular sieves (pyridine) or stored over a potassium mirror before use (diethyl ether, toluene). Hexane was dried by refluxing over potassium and stored over a potassium mirror, then degassed before use. 18-crown-6 was dissolved in DCM and dried over 4 ร
molecular sieves before being filtered and the removal of solvent in vacuo. For NMR spectroscopy, C6D6 was dried by refluxing over K, and was vacuum transferred and degassed by three freeze-pump-thaw cycles before use. NMR spectra were recorded on a Bruker AVIII HD 400 spectrometer operating at 400.07 |
668d80145101a2ffa8e358c3 | 16 | Powder XRD data were obtained on small batches of microcrystalline 2-Yb that were suspended in Fomblin ยฎ oil to prevent sample decomposition from oxygen and water. These samples were mounted on a Micromount TM and placed on a goniometer head under a cryostream to cool the sample to 100 K, freezing the Fomblin ยฎ to suspend the crystallites for the duration of the experiment. The PXRD data were measured on a Rigaku FR-X diffractometer, operating in powder diffraction mode using Cu Kฮฑ radiation (ฮป = 1.5418 ร
) with a Hypix-6000HE detector and an Oxford Cryosystems nitrogen flow gas system. Data were collected between 3-70 ยฐฮธ, with a detector distance of 150 mm and a beam divergence of 1.0 mRad. 78 Xray data were collected using CrysAlisPro. For data processing the instrument was calibrated using LaB6 as standard. Then, X-ray data were reduced and integrated using CrysAlisPro. Steady state and time resolved emission spectra were recorded at room temperature from 300-800 Continuous Wave (CW) X-band (ca. 9.4 GHz) spectra were recorded with a Bruker EMX spectrometer fitted with a Super High Q X-band resonator and at Q-band (ca. 33.9 GHz) microwave frequency using a Bruker EMX300 spectrometer. Polycrystalline and frozen solution (9 : 1 toluene : hexane, 10mM) samples of 1-Eu, 3-Eu, and 4-Eu were sealed in quartz X-band and Q-band EPR tubes in vacuo; |
668d80145101a2ffa8e358c3 | 17 | samples were lightly ground with a mortar and pestle to reduce the amount of sample decomposition, but we note that some effects due to polycrystallinity remain in the spectra below. The presence of a very sharp resonance at g = 2.00 is attributed to an impurity in the quartz EPR tubes, and serves as an internal reference for comparing relative intensities. The spectra were simulated, using EasySpin 6.0.0-dev.48 using the pepper function. The OpenMolcas 71 (version v19.11-d14be45) quantum chemistry package was used to perform CASSCF-SO calculations on 1-Ln, 3-Sm, 3-Eu, 4-Sm and 4-Eu to determine the electronic structures. The molecular geometry from a single crystal XRD structure was used by selecting a single molecule from the asymmetric unit and taking the largest disorder component only. Electron integrals were calculated using basis sets from the ANO-RCC library with VTZP quality on the metal atom, VDZP quality on the P and O atoms and VDZ quality on all other atoms, employing the second-order DKH Hamiltonian for scalar relativistic effects. Resolution of identity Cholesky decomposition (RICD) of the two-electron integrals with atomic compact Cholesky decomposition (acCD) auxiliary basis sets was employed to reduce computational demand. The molecular orbitals (MOs) for 1-Ln, 3-Sm, 3-Eu, 4-Sm and 4-Eu were optimized using state-averaged CASSCF (SA-CASSCF) wave functions; the active space taken was the valence 4f electrons and seven 4f orbitals. For 1-Eu, one Eu(II) ion was replaced with Sr(II), a diamagnetic equivalent, to determine the electronic structure and magnetic properties of one spin center at a time. The SA-CASSCF wavefunctions included 1 octet, 48 sextets, 392 quartets, and 560 doublets; a subset of these |
668d80145101a2ffa8e358c3 | 18 | roots (1 octet, 48 sextets, 119 quartets, 113 doublets) were then mixed by spin orbit coupling using the restricted active space self-interaction (RASSI) method. The SINGLE_ANISO module was used to decompose the resulting spin-orbit wave functions into the CF Hamiltonian formalism. Diamond was employed for molecular graphics. Gaussian 16 Rev C.01 84 was used to perform density functional theory calculations on 3-Yb and 4-Yb. Hydrogen positions were optimized while freezing all heavier atoms at the crystal structure positions using the PBE density functional with the Stuttgart RSC 1997 ECP for Yb and cc-pDVZ basis set for the ligand atoms. Dispersion interactions were treated using Grimme's D3 dispersion correction. Topological analysis of the total electron density was performed using Multiwfn 3.8, where the electron density at the bond critical point and delocalization indices were calculated to quantify the Yb-P interaction strength. |
668d80145101a2ffa8e358c3 | 19 | To a precooled (-78 ยฐC) suspension of [SmI2(THF)2] (0.5484 g, 1 mmol) in diethyl ether (10 mL), a suspension of KPโฒโฒ (0.6494 g, 3 mmol) in diethyl ether (10 mL) was added dropwise. The resultant green reaction mixture was stirred for 1 hour at -78 ยฐC before being allowed to warm to room temperature over 20 mins. All volatiles were removed in vacuo, the green solid was extracted with pentanes (30 mL) and was filtered to yield a dark green solution. The filtrate was concentrated slowly to ca. |
64712636be16ad5c57f82221 | 0 | Mitigating greenhouse gas emissions in the energy and chemical industries requires a transition from current fossil fuel-based thermochemical processes to carbon emission-free electrochemical processes. Given its abundance, atmospheric molecular oxygen is an attractive oxidant for the generation of electricity from chemical fuels (such as H2 or methanol in fuel cells or metals like Zn, Fe, or Al in metal-air batteries ) or for the distributed on-site electrochemical production of hydrogen peroxide. In aqueous systems at room temperature, these separate applications require different selectivity in the terminal product of oxygen reduction, where generating hydrogen peroxide terminates at a net 2-electron/2-proton reduction (hydrogen peroxide production reaction, HPPR: O2 + 2H + + 2e -โ H2O2, E 0 = 0.695 V vs. SHE) and fuel cell or metal-air battery applications ideally reduce oxygen to water through a net 4-electron/4-proton process (oxygen reduction reaction, ORR: O2 + 4H + + 4e -โ 2H2O, E 0 = 1.23 V vs. SHE). |
64712636be16ad5c57f82221 | 1 | Although not universal, it is generally assumed that an electrocatalyst is needed to reduce oxygen. Recently, organic mixed ionic-electronic conducting polymers (OMIECs) have been reported to exhibit catalytic behavior for electrochemical reactions. OMIECs are semiconducting conjugated polymers with synthetically tunable structures and transport properties enabled by electrochemical ion-insertion redox reactions. These electrochemical doping reactions drive the insertion of mobile ions and electrons (or holes), providing functionality for a range of electrochemical devices. For electrocatalysis, hole/anion transporting OMIECs (p-type), including PEDOT (poly(3,4-ethylenedioxythiophene)) variations, have been studied for oxygen reduction and display selectivity for the 2-electron reaction. BBL (poly(benzimidazobenzophenanthroline)), an electron/cation transporting (n-type) OMIEC, has also been shown to be active for the ORR, with pH-dependent selectivity for the 4-electron versus the 2-electron product. These studies highlight the potential of OMIECs as electrodes for oxygen reduction, but the fundamental design principles that control their activities and selectivity are still being developed. |
64712636be16ad5c57f82221 | 2 | Our work sets out to identify the operative oxygen reduction mechanism of OMIECs through a multi-faceted experimental and theoretical approach. We primarily focus our efforts on p(NDI-T2 P75), a naphthalene-1,4,5,8-tetracarboxylic-diimide-bithiophene copolymer with 75% polar side chains (Figure ). NDI-T2 copolymers are acceptor-donor copolymers and are known to have good charge transport properties relative to other n-type OMIECs. Notably, this OMIEC is stable in a broad pH range, is structurally similar to a previously described oxygen-reducing OMIEC, and has heteroatoms and functional groups reported to be vital for electrocatalysis on carbonbased materials, including N, carbonyl groups (C=O), and S. Figure . Structure of p(NDI-T2 P75), an acceptor-donor random copolymer with a backbone composed of an electron-deficient naphthalene diimide (NDI) unit and an electron-rich bithiophene (T2) unit. The hydrophilic side chains based on ethylene glycol, R , attached to 75% of monomer units (P75) facilitate solvated ion-insertion, and the hydrophobic side chains based on alkyl chains, R 2 , provide improved electrochemical stability. We combine rotating ring-disk electrochemistry, operando spectroscopy techniques, and ab initio and steady-state microkinetic simulations to evaluate the performance and mechanism of p(NDI-T2 P75) for oxygen reduction. We validate the generality of our mechanistic conclusions by benchmarking the performance of an expanded selection of p-and n-type OMIECs. The rigorous approach to studying electrochemical H2O2 production presented in this paper can be extended to materials beyond OMIECs and facilitates a guided exploration to understand the operating principles of oxygen reducing electrodes. |
64712636be16ad5c57f82221 | 3 | Prior to all electrochemical experiments, p(NDI-T2 P75) was purified to remove residual Pd contamination from polymer synthesis (Figure , Table ), resulting in low Pd concentrations of 2 -5 ppm. While it has been reported that a polymer much like p(NDI-T2 P75) primarily produces H2O upon reducing O2 in a 4-electron reaction, we demonstrate that the removal of residual Pd is crucial for evaluating intrinsic performance. We find residual Pd impurities result in a positive shift in the measured halfwave voltage, E1/2, for oxygen reduction and modifies the selectivity from H2O2 towards H2O as the majority product (Figure ). We use rotating ring-disk electrochemistry (RRDE) to investigate the behavior of p(NDI-T2 P75) in [K + ] = 1 M aqueous electrolytes with pH varied between 7.0 and 14.2. Characteristic voltammograms at a scan rate of 5 mV s -1 for p(NDI-T2 P75) in Ar-saturated electrolytes are shown in Figure . For all tested pH values, three reversible redox peaks are observed. The first two peaks, which appear at more positive voltages, correspond to electron polaron formation and the third peak corresponds to the formation of the electron bipolaron. We confirm the n-type nature of these electron polaron formation reactions through conductivity measurements using an interdigitated electrode array in an organic electrochemical transistor (OECT) architecture (Figure ). The formation of the electron polaronic state is associated with electron/cation doping of the polymer film resulting in an electronic conductivity enhancement by a factor of >10 . We hypothesize that the voltage offset for the first two electron polaron peaks arises from contributions of the mixed side chain composition and the influence of mesoscopic domains of ordered (crystalline) and disordered (amorphous) aggregates in the polymer film. It has been demonstrated that inter-chain electronic state hybridization can occur in ordered domains, which stabilizes the LUMO state (most positive reduction voltage) similar to the effect of charge-transfer hybridization, potentially resulting in a shift of the observed reduction voltage. Integration of the reduction current suggests ~50% of the film can be reduced to the bipolaronic state at 5 mV s -1 (corresponding to a charging time of 4 minutes), confirming fast charge transport through the bulk of the polymer electrode. Figure . a) Ar-saturated reduction/oxidation voltammograms at a scan rate of 5 mV s -1 . b) O2saturated RRDE measurement at 1600 r.p.m. with a scan rate of 0.5 mV s -1 , showing a high ring current (Ering = 1.20 V vs. RHE). c) Percent H2O2 produced calculated using the ring current from (b) and the electrode collection efficiency (N = 0.25). d) Number of electrons transferred during the ORR as calculated through Kouteckรฝ-Levich analysis. e) Quantification of Faradaic efficiency obtained from constant current electrolysis (-2 mA cm -2 ) in an H-cell performed on a carbon fiber paper supported p(NDI-T2 P75) film in 0.1 M KOH. |
64712636be16ad5c57f82221 | 4 | Interestingly, the limiting current density for oxygen reduction on p(NDI-T2 P75) remains constant regardless of polymer mass loading on the disk (Figure ), suggesting that oxygen reduction is limited to the electrode surface. Additionally, the observed limiting current is half the limiting current of a polycrystalline platinum electrode (Figures ) tested in identical (%) |
64712636be16ad5c57f82221 | 5 | conditions. Assuming Pt reduces O2 to H2O through a net 4-electron process (as evidenced through the absence of ring-current and Kouteckรฝ-Levich analysis in Figure ), the limiting current density measured for p(NDI-T2 P75) suggests the electrode is highly selective for the 2electron product. High Faradaic efficiencies for H2O2 were additionally observed through collection on the ring electrode (Figures ), Kouteckรฝ-Levich analysis (Figures ), and constant current electrolysis performed in an H-cell (Figure ). The polymer exhibited stable performance at -2 mA cm -2 for two hours in the H-cell measurements (Figure ) with negligible changes in chemical structure observed through post-mortem 1 H NMR analysis (Figure ). We observe a decrease in limiting current densities with increasing pH. In accordance with observations by McCreery et al., the pH-dependent variation of limiting current density suggests the reaction proceeds through a single-electron transfer reduction of dioxygen to superoxide followed by disproportionation of superoxide into hydro(gen) peroxide and dioxygen. This is explained in detail in Section 7 of the Supplementary Information. |
64712636be16ad5c57f82221 | 6 | Comparing the trends in bulk redox and oxygen reduction, we observe an ~59 mV pH -1 positive shift in voltage versus RHE for both the electron polaron/bipolaron reactions and oxygen reduction (Figure ). Although the overall ORR and HPPR are both proton-coupled, the observed pH-dependence reflects that both the bulk ion-insertion reactions and the rate-limiting step of oxygen reduction do not involve proton transfer. Instead, charge compensation for the electron polaron/bipolaron reaction is likely provided by K + , the cation with majority concentration in these experiments. |
64712636be16ad5c57f82221 | 7 | The UV-Vis spectra of p(NDI-T2 P75) during reduction from its neutral to its bipolaron state in Ar-saturated 0.1 M KOH is presented in Figure . The increase in optical density at ~500 nm and the respective decrease near 720 nm correspond to electron polaron formation, and the increase in optical density near 630 nm is indicative of the electron bipolaron state. Figure presents the differential optical density of the electron polaron feature at 500 nm with respect to time and voltage. The appearance and disappearance of the electron polaron feature in Arsaturated electrolyte correspond well with the voltages observed in cyclic voltammetry. In O2-saturated electrolytes, however, the intensity of the electron polaron feature was notably suppressed and shifted to more cathodic voltages. These results suggest the rate of O2 reduction exceeds the rate of electron polaron formation in the film, where oxygen present in the electrolyte can rapidly consume the energetic electrons of the electron polaronic state of p(NDI-T2 P75) to generate hydro(gen) peroxide. Further evidence of this process is provided in Supplementary Figure in which we track the polaron feature at 500 nm as the electrolyte feed to the spectroelectrochemical flow cell was switched repeatedly between Ar-saturated and O2-saturated 0.1 M KCl electrolytes, resulting in the appearance and disappearance of the polaronic state, respectively. neutral, polaronic, and bipolaronic states of the polymer measured at characteristic voltages in Ar-saturated electrolyte, shown in Figure . |
64712636be16ad5c57f82221 | 8 | Starting with the initial reference spectrum obtained at the open-circuit voltage, the symmetric C=O stretching band at 1708 cm -1 can be used as an indicator for neutral repeat units as it has been shown to disappear as NDI-T2 becomes highly reduced. As the applied voltage reaches 0.4 V vs. RHE and the material is reduced to its singly-reduced, electron polaron state, two new bands at 1538 and 1370 cm -1 arise. These bands can be ascribed to a shift of the pristine bands at 1610 and 1572 cm -1 , associated with the C-C stretching in the NDI unit. The intensities of these vibrations increase with applied voltage and are assigned to electron polaron formation. Moreover, the intensity ratio of the triplet bands between 1496 and 1400 cm -1 undergoes a few changes, with a substantial drop of intensity of the bands at 1434 and 1406 cm -1 , assigned to collective vibrational displacements of the C-C stretching of the NDI and the T2 units. When the applied voltage reaches 0.2 V vs. RHE, the bands assigned to the electron polaron at 1583 cm -1 and 1370 cm -1 gain further intensity, while the triplet between 1496 and 1400 cm -1 broadens and loses intensity. Additionally, at this voltage, the band at 1708 cm -1 , which can be used to indicate the presence of neutral NDI-T2 repeat units, disappears. Once -0.1 V vs. RHE is applied, new vibrational features arise that have not previously been reported, namely the bands at 1344 and 1137 cm -1 , which, based on the associated electrochemical features, are likely associated with the formation of the electron bipolaron. |
64712636be16ad5c57f82221 | 9 | When similar measurements are performed in O2-saturated electrolyte, we observe a suppression of the electron polaron features until the applied voltage reaches -0.1 V vs. RHE (in comparison to 0.4 and 0.2 V vs. RHE in Ar-saturated electrolyte). Additionally, bipolaronic vibrational bands are not observed in the presence of oxygen, and the neutral band in the Raman spectra does not disappear. These results confirm that electron polaron formation is suppressed through rapid consumption by O2. Notably, no new features were observed in O2saturated electrolyte that could be explained through a surface bound oxygen intermediate. O-O stretching modes for superoxide adsorbates on metals have been reported in the range of ~1000 -1200 cm -1 , which overlaps with one of the neutral p(NDI-T2 P75) Raman modes and, as such, cannot be unequivocally ruled out. However, there is no apparent intensity increase in this region during ORR. Surface bound peroxides have stretching modes at lower wavenumbers, from 640 -900 cm -1 , and we observe no increased Raman intensity in this region. Similar operando spectroscopic results can be observed in 0.1 M KCl (Figure ). |
64712636be16ad5c57f82221 | 10 | While our results reveal p(NDI-T2 P75) reacts directly with oxygen to form hydro(gen) peroxide, our operando spectroscopy results do not demonstrate evidence for a surface-bound intermediate expected for a catalytic, or "inner-sphere," pathway. In this regard, we briefly describe the energetic contributions to ORR reactivity, highlighting the non-catalytic outersphere pathways for oxygen reduction which may occur at low overpotentials at high pH. Figure presents a general, though not necessarily universal, framework to understand oxygen reduction to the 2-electron product in alkaline electrolytes (pH > 11.7) involving the commonly hypothesized surface adsorbate and solvated intermediates. (Electro)catalysis occurs through the adsorption of reaction intermediates (chemisorption/surface bond formation), which modulates their reduction voltages by -ฮGads/F, where ฮGads is the Gibbs energy of adsorption and F is Faraday's constant. In the Alternative non-catalytic pathways for the 2-electron oxygen reduction, however, may occur without adsorption. In these outer-sphere pathways, the first electron transfer reaction to oxygen is crucial. This initial reduction reaction is generally considered the rate-limiting step (RLS) due to its very negative standard reduction voltage. At pH > 4.88 (pKa of HO2), this step yields the superoxide radical, O2 โข-, as shown in Reaction [1]: |
64712636be16ad5c57f82221 | 11 | The standard reduction voltage for this reaction is ๐ธ ๐ 2 ๐ 2 โข- โ 0 = -0.33 V vs. SHE. However, this voltage corresponds to a superoxide activity of 1. In contrast, its formal reduction voltage, E 0 ', can be calculated as follows, where R is the ideal gas constant, T is the absolute temperature, ๐ ๐ 2 is the partial pressure of O2, and a is the activity of species in the aqueous solution: |
64712636be16ad5c57f82221 | 12 | The low stability of superoxide suggests that its concentration is unlikely to exceed the concentration of dissolved O2 in the electrolyte from which it is produced. Although O2 solubility is pH-dependent, we can estimate the superoxide and oxygen activity through Henry's Law where the Henry's Law Constant for O2 in water is ๐ป ๐ 2 ๐๐ = 1.3 x 10 -3 M atm -1 : |
64712636be16ad5c57f82221 | 13 | Using this value for ๐ ๐ 2 โข-we obtain a formal reduction voltage of ๐ธ ๐ 2 ๐ 2 โข- โ 0โฒ = -0.16 V vs. SHE, which defines the reversible voltage of dissolved oxygen reduction to superoxide in aqueous electrolytes. We note this value corresponds well with commonly observed halfwave voltages (E1/2) on carbon-based electrodes in aqueous electrolytes. In a fully outer-sphere reaction pathway, the second electron reduction yields hydrogen peroxide, H2O2, at pH โค pKa = 11.7, or hydroperoxide, HO2 -, at pH > pKa and may proceed through the following reactions depending on pH: |
64712636be16ad5c57f82221 | 14 | Previous work has demonstrated that this reaction proceeds through a hydroperoxyl, HO2, intermediate resulting in pH-dependent kinetics. For the homogeneous disproportionation of radiolytically produced O2 sphere 1-electron reduction to form superoxide followed by disproportionation. We summarize the energetics involved in the catalytic (inner-sphere) and non-catalytic (outersphere) pathways through a Pourbaix diagram (modified for the RHE scale) for the reactive oxygen species in Figure . When the ORR halfwave voltage (E1/2) for oxygen reduction is |
64712636be16ad5c57f82221 | 15 | ), it can be assumed the reaction proceeds through an innersphere catalytic pathway with at least partial selectivity for H2O production. This means that the electron transfer number (n) in this regime will be greater than two because reducing O2 to H2O requires four electrons. When the E1/2 for oxygen reduction is below |
64712636be16ad5c57f82221 | 16 | ), there are other factors limiting oxygen reduction, namely conductivity. Still, in this case there is no requirement for catalytic adsorption and oxygen reduction may proceed through the outer-sphere mechanism. and UV-Vis spectroscopy revealed that there is a direct chemical reaction between the polaronic state of the polymer and oxygen, forming a reduced oxygen species and re-oxidizing the polymer to the neutral state in the process (i.e. an electron transfer process mediated by the polymer). To rationalize these observations, we investigate the O2 adsorption energetics of p(NDI-T2 P75) at different redox states and potential binding sites (Figure ) using density functional theory (DFT) to. A full description of the DFT calculation methods is provided in the Supplementary Information. |
64712636be16ad5c57f82221 | 17 | The neutral state of the polymer showed no affinity for binding O2 (Figure ). For the reduced states, the reduced complex (NDI-T2 + O2) -and the case where isolated O2 bears the entire negative charge are energetically identical and favorable relative to the case where the NDI-T2 alone bears the charge (Figure ). Inspection of the partial charge distribution for the reduced complex (Figure ) shows that the extra -1 e -charge is mainly localized on the O2 molecule, leaving the NDI-T2 monomer almost neutral, irrespective of the O2 location in the complex (Figure ). These results suggest that the lowest energy configuration is achieved when the added negative charge resides on O2 and that the charged O2 species (superoxide) is not stabilized by interaction with NDI-T2. Thus although there is electron transfer between the reduced polymer and O2 to form superoxide, it appears this process is chemical in nature and follows an outer-sphere process. |
64712636be16ad5c57f82221 | 18 | for p(NDI-T2 P75) at all pH's explore in this study (suggesting superoxide formation is rate-limiting), if oxygen reduction on p(NDI-T2 P75) followed a serial 1+1 outer-sphere electron transfer pathway, we would expect a constant limiting current across the pH range (assuming that O2 solubility remains approximately the same). However, in addition to observing a pH dependence in the E1/2 for oxygen reduction in the RRDE experiments, we observe a decrease in the limiting current density with increasing pH. This observation is unlikely to be due to surface roughness as the limiting current is not dependent on the mass of p(NDI-T2 P75) on the electrode (Figure ). As demonstrated previously for metal-free carbon electrodes by Yang and McCreery, this observation can be explained through disproportionation of the superoxide intermediate to form O2 and hydro(gen) peroxide. The re-formed O2 can then be electrochemically reduced to superoxide generating a limiting current density that falls between the expected values from the Levich equation for an n = 1 and n = 2 electron transfer reaction, depending on the disproportionation rate. The disproportionation rate has been observed to scale with the concentration of hydroperoxyl, HO2, which decreases with increasing pH and thus lowers the observed limiting current density. We explore the possibility of disproportionation of superoxide following its formation from the outer-sphere reaction between the electron polaron in p(NDI-T2 P75) and oxygen through the use of a microkinetic simulation based on the following EC'D mechanism: |
64712636be16ad5c57f82221 | 19 | In this mechanism, the first electron transfer (E) converts the neutral polymer, N, into its reduced electron polaron state, P. The now electronically conductive polaronic state of the polymer can chemically reduce oxygen generating superoxide, O2 โข-, while regenerating the neutral state of the polymer (C'). In the next step, superoxide radicals then disproportionate into O2 and hydro(gen) peroxide (D) (Figure ). We note that Step [8] corresponds to the elementary reaction considered in the DFT calculations. |
64712636be16ad5c57f82221 | 20 | A full description of the microkinetic model, solution procedure, and error analysis is included in the Supplementary Information. In short, reactions [7-9] are cast into a set of rate expressions, where ๐ ๐ * is an electrochemical rate coefficient (dependent on applied voltage), ๐ ๐ is a chemical rate coefficient (independent of applied voltage), ๐ ๐ and ๐ ๐ , are the surface fractions of the polymer in its neutral and electron polaron states, respectively, and ๐ถ ๐ ๐ is the concentration of species a at the surface of the electrode: |
64712636be16ad5c57f82221 | 21 | The fraction of the neutral and electron polaron states of the polymer as a function of applied voltage are determined using Butler-Volmer kinetics and the steady state approximation. The steady state solutions are simultaneously fit to both the experimental log current densityvoltage (or Tafel) data and the linear disk current density-voltage RRDE data at all pH's to extract a self-consistent set of fitting parameters describing the thermodynamics and kinetics of each elementary reaction step. The pH-independent fitting parameters are the formal reduction voltage of reaction [7]: ๐ธ 1 0โฒ ; and the associated chemical and electrochemical reaction rate constants and symmetry coefficients: ๐ 1 0 , ๐ฝ 1 , ๐ ๐,2 . The ๐ ๐,3 parameter is treated as pHdependent, and each experimental pH has its own ๐ ๐,3 fit. Note that the equilibrium constant for reaction [8], ๐พ 2 , is defined through |
64712636be16ad5c57f82221 | 22 | The results of the simulation are shown in Figure , and the values of the fitting parameters are presented in Table . As is evident in Figure and, the simulation provides excellent fits of the experimental data at all pH using a self-consistent set of fitting parameters and can explain the observed trends in pH-dependent limiting current and Tafel behavior across 4-orders of magnitude in kinetic current. The formation of the electron polaron in Step 1 of the reaction mechanism is identified as the RLS, which helps to explain the experimental observation of polaron suppression, whereby oxygen rapidly oxidizes any polaron formed through the chemical reaction in Step 2. We also note that, assuming a symmetry coefficient of ฮฒ = 0.5 -0.6 for the outer-sphere electron transfer in Reaction [8], the observed rate coefficient of ๐ ๐,2 = 12.1 cm s -1 would have a standard electrochemical rate constant of ๐ ๐ 2 /๐ 2 โข- 0 = 0.003 -0.011 cm s -1 which agrees very well with the observed rate constant of ๐ ๐ 2 /๐ 2 โข- 0 = 0.003 ยฑ 0.005 cm s -1 for oxygen reduction to superoxide reported by McCreery et al. In Figure observed in this work compared to results reported by Bielski and Allen, showing the expected negative linear dependence but a shallower slope, suggesting the disproportionation process in this work is catalyzed. |
64712636be16ad5c57f82221 | 23 | We further examine our proposed EC'D mechanism by benchmarking the oxygen reduction performance of select metal-free OMIECs as well as glassy carbon, which is the underlying substrate for these measurements and reduces oxygen through an outer-sphere pathway (Figure ). The polymers chosen were p(NDI-T) , p(gPyDPP-MeOT2) , and p(g3T2) (Figure ). Fit E 0 1 First, we employ p(NDI-T) , which can be reduced to its electron polaronic state at voltages more positive than ๐ธ ๐ 2 ๐ 2 โข- โ 0โฒ |
64712636be16ad5c57f82221 | 24 | . Following this rationale, it is not surprising that p(NDI-T) and p(NDI-T2 P75) exhibit comparable HPPR activity (Figure ). Here, however, it is unclear whether polaron formation remains rate-limiting and the observed increase in current relative to p(NDI-T2 P75) may be due to a larger capacity of reactive electrons in the film or from a difference in electrochemically active surface area. |
64712636be16ad5c57f82221 | 25 | , the E1/2 for oxygen reduction will correspond to the voltage of this first redox reaction, which is consistent with the proposed outer-sphere electron transfer for the initial reduction of O2. Here, we take the E1/2 for oxygen reduction to correspond to the largest increase in reduction current (~0.17 V vs. RHE), though we note the initial activity in p(gPyDPP-MeOT2) is likely caused by exposure of the underlying glassy carbon substrate. In agreement with this finding regarding the crucial nature of the first electrochemical step, p(g3T2), a p-type OMIEC, does not effectively reduce oxygen through the EC'D mechanism with high activity because of the lack of ion-insertion redox at appropriate reduction voltages (Figure )-although here too there is likely some exposure of the underlying glassy carbon substrate. |
64712636be16ad5c57f82221 | 26 | The effects of polaron formation voltage, ๐ธ 1 0โฒ , on oxygen reduction were additionally captured using the previously discussed microkinetic model (Figure ). As ๐ธ 1 0โฒ shifts to negative voltages, the E1/2 for oxygen reduction shifts negative accordingly. Inversely, for polaron formation voltage greater than p(NDI-T2 P75) (๐ธ 1 0โฒ > -0.31 V vs. SHE), the halfwave voltage for oxygen reduction is pinned by the outer-sphere formal reduction potential for the O2/O2 |
64712636be16ad5c57f82221 | 27 | Figure demonstrates that the disproportionation rate, ๐ ๐,3 , primarily affects the limiting current densities but not the halfwave voltage of oxygen reduction, where low disproportionation rates result in only a single electron transfer (corresponding to exclusively superoxide formation) while high disproportionation rates push the electron transfer number towards n = 2 due to the feedback loop of generated O2. Thus, the catalytic disproportionation of superoxide helps to rationalize the difference in observed oxygen reduction current between the p(NDI-T2 P75) coated electrode and the bare glassy carbon electrode-which shares the same E1/2 but only generates half of the limiting current in 0.1 M KOH (Figure ). On glassy carbon, which is apparently less catalytic for superoxide disproportionation, the limiting current density corresponds to a single electron transfer to form superoxide which builds up at the electrode surface excluding O2 resulting in a decrease in current density with increasing potential. While the side chains for the tested OMIECs are not identical, they all contain a polar (hydrophilic) portion, which is essential for bulk ion-insertion redox in aqueous electrolytes. In contrast, an NDI-T2 polymer with entirely nonpolar side chains was found to be incapable of electrochemical activation and oxygen reduction in aqueous electrolytes (Figure ). The |
64712636be16ad5c57f82221 | 28 | OMIECs studied here span polymers with variations in backbone chemistry (homo-and copolymers) and side chain chemistry, suggesting that the reactivity of many polymers within this class of materials may be rationalized through this EC'D mechanism when the terminal oxygen reduction product is hydro(gen) peroxide. However, we note our sample set of materials is still too narrow to make general claims for the whole range of OMIECs, and there are likely opportunities for alternative pathways that may even result in the 4-electron water product. |
64712636be16ad5c57f82221 | 29 | In this work, we investigated a set of organic mixed ionic-electronic conducting polymers as single-phase electrodes for the oxygen reduction reaction, focusing on p(NDI-T2 P75) as a model system. In p(NDI-T2 P75), the halfwave voltages (E1/2) for the bulk ion-insertion redox reactions (electron polaron formation) and oxygen reduction were found to be independent of pH, and the dominant product was found to be hydro ). For polymers with ๐ธ 1 0โฒ < ๐ธ ๐ 2 /๐ |
64712636be16ad5c57f82221 | 30 | may bypass the chemical step where electrochemical oxygen reduction to superoxide then becomes rate-limiting. Additionally, this interpretation explains the difference between the observed performance of the glassy carbon substrate and the mixed conducting polymers through the difference between their superoxide disproportionation rates, ๐ ๐,3 . As seen on p(NDI-T2 P75), the catalytic disproportionation of superoxide leads to the same observed E1/2 for oxygen reduction but results in almost twice the limiting current density. |
64712636be16ad5c57f82221 | 31 | While our results suggest the polymers in this study do not function as electrocatalysts for the initial reduction of oxygen to superoxide (i.e. through an adsorption-driven inner-sphere process), they suggest there may be opportunities for polymers to serve as catalysts for the disproportionation of superoxide into O2 and HxO2 . Additionally, the lack of metal centers that would likely drive intermediate adsorption and operation through the outer-sphere pathway enables p(NDI-T2 P75) to highly selective for H2O2 production at low overpotentials when operating in high pH, suggesting promising opportunities for this emerging class of polymer electrodes. |
64712636be16ad5c57f82221 | 32 | Polymer solutions of p(NDI-T2 P75) were prepared by dissolving the polymer in chloroform at a mass concentration of 11.7 mg/mL. Prior to deposition, the RRDE was cleaned and polished with Al2O3 (particle size 0.05 ฮผm) to a mirror finish. To create thick polymer films (estimated thickness = 1 ฮผm), 10 ฮผL of this solution was dropcast onto the glassy carbon disk of a rotating ring disk electrode (RRDE) (E6R1, Pt ring, polyether ether ketone (PEEK) shrouded, Pine Research) and the electrode was rotated at 400 r.p.m. for 2 min. |
64712636be16ad5c57f82221 | 33 | Three-electrode electrochemical measurements were conducted in an alkaline-resistant PTFE cell (AF01CKT1001, Pine Research Instrumentation) using a VSP-300 potentiostat (Biologic), with a leakless Ag/AgCl reference electrode (ET069, eDAQ), a coiled Pt wire counter electrode, and the RRDE with the dropcast or spin cast fims p(NDI-T2 P75) film. All voltages are referenced versus the reversible hydrogen electrode (RHE), where the RHE voltage of the Ag/AgCl reference for each experiment was standardized against a bulk RHE electrode (Hydroflex Hydrogen Reference Electrode, eDAQ) in each electrolyte prior to testing. The 0.1 M KOH electrolytes were prepared using KOH and ultrapure de-ionized water ("DI water", 18.2 Mฮฉ resistance, MilliQ). The electrolytes for the pH series were made by mixing varying ratios of 1 M KHCO3 and 1 M KOH stock solutions, where each stock solution was prepared in advance with ultrapure de-ionized water. Prior to electrochemical characterization, Ar gas was allowed to bubble through the solution for at least 30 minutes. For each experiment, the uncompensated resistance, R, was measured at open circuit voltage (polymer is in the pristine, neutral state) through electrochemical impedance spectroscopy (EIS) at the high frequency intercept and iR correction was completed manually during the data analysis phase. In the 1 M [K + ] electrolytes, The R values in the 1 M [K + ] 0.1 M [K + ] electrolytes were about 15 ฮฉ and 58 ฮฉ, respectively. In Ar-saturated solution, the Pt ring was electrochemically cleaned according to the procedure outlined by Chen et al. The p(NDI-T2 P75) film was then cycled five times between 0.2 V and -0.5 V vs. Ag/AgCl and another five times between 0.2 V and -1.0 V vs. Ag/AgCl at 5 mV s -1 scan rate. O2 gas was then bubbled through solution for at least 30 minutes to saturate the electrolyte. In O2-saturated solution, RRDE measurements were completed at different rotation rates (400, 900, 1200, and 1600 r.p.m., MSR Rotator, Pine Research) while the p(NDI-T2 P75) was cycled between 0.2 V and -1.0 V at 5 mV s -1 and the ring was held constant at 1.2 V vs. RHE. At 1600 r.p.m., an extra RRDE measurement was completed following the same procedure as before, with a changed scan rate of 0.5 mV s -1 . All the measured disk and ring currents were normalized by the disk electrode area (~0.196 cm 2 ). |
64712636be16ad5c57f82221 | 34 | Electrolysis was performed in a 2-compartment glass H-cell with a ceramic separator. The cathode compartment housed a 0.25 cm 2 Teflon-coated carbon paper (AVCarb P50T) electrode loaded with ~4 mg cm -2 polymer, and a leakless Ag/AgCl reference electrode (eDAQ). In the anode compartment, Ni foam (MTI) was used as the counter electrode, and each compartment was filled with 5 mL 0.1 M KOH. Electrolytes were stirred and O2 was bubbled into the cathode test compartment throughout the experiment. At 20-minute intervals 600 ฮผL aliquots of sample were collected from the cathode compartment and replaced with fresh electrolyte. |
64712636be16ad5c57f82221 | 35 | Peroxide concentration was measured using a Cerium(IV) sulfate colorimetric assay. Samples were prepared by mixing 175 ฮผL of analyte, 700 ฮผL of 1 mM Ce(IV) solution (in 0.5 M H2SO4), and 875 ฮผL of 0.5 M H2SO4. The UV-visible spectrum (Cary 6000i UV-Vis Spectrometer) was recorded in transmission mode, and the intensity at 318 nm was used to quantify the Ce(IV) concentration in the sample, from which the H2O2 concentration was determined. A linear calibration curve was constructed by measuring analytes of known H2O2 concentration. When necessary, analytes were diluted in 0.5 M sulfuric acid to be within the measurement range of the assay. |
64712636be16ad5c57f82221 | 36 | Thin films on ITO coated glass substrates were prepared by spin coating. 65 ฮผL of a p(NDI-T2 P75) solution (10 mg/mL) were spin coated onto a cleaned ITO coated glass substrates at 1000 r.p.m.. A thin layer of epoxy glue was applied at the edges of the film to prevent delamination (5 minute setting, Loctite Epoxy). The p(NDI-T2 P75) film was then placed into a quartz cuvette (DLC-300-Q-20, Starna Cells) filled with 0.1 M KOH. The electrochemical measurements were completed in a three-electrode set-up with a PEEK-shrouded leakless Ag/AgCl reference (eDAQ) and a coiled Pt wire as the counter. The UV-vis measurements were completed with a Tungsten light source (HL-2000-LL, Ocean Optics) and a spectrometer (QE Pro, Ocean Optics) and the instruments were synchronized using a Matlab code. |
64712636be16ad5c57f82221 | 37 | In-situ Raman spectra were acquired by a self-design electrochemical cell. (3D-printed by Protolabs, UK) with a miniature leakless Ag/AgCl reference electrode (eDAQ), and platinum sheet counter electrode (Sigma-Aldrich, USA). The samples were spin coated at 1000 r.p.m. for 60 s on an ITO slide (Kintec Inc, Hong Kong). Spectra were acquired in the Raman Stokes region using 532 nm excitation wavelength, in the range 150-2200 cm -1 (Horiba Xplora Confocal Raman). A 10x Olympus objective was used to perform measurements with laser power set at < 0.963 mW to avoid sample degradation. For a good signal-to-noise ratio, each spectrum was taken over an integration time of 20-30 s and averaged over three measurements. The spectrometer was calibrated through a Si sample using the spectral line 520.7 cm -1 . Spectra were analyzed using OMNIC software. Electrochemical signals were applied through a Bio-Logic potentiostat. A constant voltage was held for at least 30 seconds before each Raman measurement. |
60c75562842e654f73db4312 | 0 | Spin-crossover (SCO) systems are molecules or materials that can switch between two alternative spin-states, thus exhibiting switching behaviour. The transition from the low-to the high-spin state can be triggered using an external stimulus, commonly temperature, and with the spin-state change there are profound changes in the physical properties of the system. This duality is very appealing from the technological point of view, because one can envision harvesting such materials for molecular level devices or spintronic applications. The SCO phenomenon, firstly reported nearly a century ago, has grown at the interface between physics and chemistry, and the number of compounds exhibiting SCO behaviour has vastly increased over the last years. |
60c75562842e654f73db4312 | 1 | Despite the large development of the field, the vast majority of systems exhibiting SCO behaviour contain an Fe II (d ) metal center, and there is an increasing interest on expanding the set of compounds with other metals and oxidation states exhibiting this behaviour. Similarly, while coordination chemistry has proven to be key in the design of ligands that generate the right splitting among the d-based molecular orbitals for the molecule to exhibit SCO, organometallic molecules usually have a larger ligand-field splitting that leads to low-spin states, and very few examples have been reported. Among the few organometallic SCO systems reported, the manganocene family ([Mn(Cp R ) 2 ]) has provided several examples of functionalized molecules exhibiting such behaviour. More relevant is the fact that the transition temperature (T 1/2 ), defined as the temperature with equal populations of both spin-states and a key physical property in SCO systems, can be modulated in such families via the R group. |
60c75562842e654f73db4312 | 2 | The interplay between steric and electronic effects in tuning T 1/2 for the alkyl substituted manganocenes of general formula [Mn(Cp n-R ) 2 ] (n = 1 to 5 and R = Me, i Pr or t Bu) has been analyzed by means of computational studies. That work showed that the ligand field around the Mn II ion can be increased by adding more electron-donor groups, such as methyl, but that bulky substituents, such as iso-propyl or tert-butyl, can have the opposite effect due to the steric hindrance that they introduce in the molecule, which pushes the Cp rings away from the metal center. However, the effect of the functionalization of the cyclopentadienyl ligand with other groups than methyl in order to control the SCO properties of the [Mn(Cp 1-R ) 2 ] family has not been explored. |
60c75562842e654f73db4312 | 3 | In this work, we used electronic structure calculations at the density functional theory (DFT) level to evaluate the effect that different R groups have over the transition temperature (T 1/2 ). By changing the R group we will show that a fine degree of tuning over T 1/2 can be achieved, allowing for the modulation of such value in a wide range of temperatures. The presented results open the door to in silico design of new metallocenes with selected SCO properties, thus providing experimental chemists a powerful tool for the rational design of new molecules with specific transition temperatures. |
60c75562842e654f73db4312 | 4 | All density functional calculations (DFT) have been carried out with Gaussian 09 (revision D) except those with the SCAN functional, which have been performed using the Q-Chem 5.0 electronic structure suite. All calculations have been converged to 10 -8 for the density matrix elements, and the corresponding vibrational analysis was done to ensure that they were minima along the potential energy surface. The fully optimized basis set from Ahlrichs and co-workers, including polarization functions, was employed for all atoms. In particular, five different basis set schemes were tested during the benchmarking process: TZV for all elements (BS1), TZVP for manganese and TZV for the rest (BS2), TZVP for all elements (BS3), QZVP for manganese and TZVP for the rest (BS4) and QZVP for all elements (BS5) (see results section and Supporting Information). To compute the transition temperatures, several post pro-cessing scripts were used (see Supporting Information). The n-electron valence perturbation theory (NEVPT2) calculations were performed with the Orca 4.0 code. In these calculations, we employed the def2TZVPP basis set, including the corresponding auxiliary basis set for the correlation and Coulomb fitting. The active space contains the 5 d-orbitals of the metal and 4 electrons, and the ab initio ligand-field theory (AILFT) approach was employed to extract the related orbitals. |
60c75562842e654f73db4312 | 5 | Previous work on the computational modeling of T 1/2 in [Mn(Cp R ) 2 ] (R = Me, i Pr or t Bu) showed that DFT methods (OPBE in particular) were able to correctly model the SCO behaviour in such systems. However, we decide to pursue a systematic benchmark of different DFT methods aiming to be as quantitative as possible towards the calculation of the transition temperature. With that goal, we choose as a benchmark model the [Mn(Cp 1-Me ) 2 ] system for its simplicity, which allows the testing of multiple functionals and basis set schemes, and also because there is structural information in gas phase (d(Mn-C) = 2.433 ร
and 2.144 ร
for high-and low-spin respectively), as well as a proper characterization of its T 1/2 (303 K), data that will allow us to properly calibrate the computational method of choice. |
60c75562842e654f73db4312 | 6 | Several DFT methods were tested for the [Mn(Cp 1-Me ) 2 ] system, including TPSSh, OPBE, OLYP, M06L, 34 B3LYP, 35 B3LYP* and SCAN. A full optimization in both high-and low-spin state using BS3 was done, followed by the corresponding vibrational analysis to ensure its minimum nature. The corresponding spin-state energy differences as well as relevant geometric parameters and the calculated T 1/2 (where meaningful) are given in table . Mn-C bond lengths compared with the experimental values as well as an appropriate energy gap for SCO to occur. Among the most promising methods, a systematic exploration of the basis set effect was carried out. A total of five different basis set schemes were tested (BS1 to BS5, see computational details). The results are summarized in Table . |
60c75562842e654f73db4312 | 7 | As can be seen from table , several combinations achieve a remarkable precision when computing the experimental transition temperature (T 1/2 = 303 K), as has been previously reported. An optimal balance between computational cost and accuracy can be achieved using OLYP/BS4. The same is observed by analyzing the metal-ligand bond lengths (see Supporting Information), as Mn-C bond lengths are in excellent agreement with the experimental data (d(Mn-C) = 2.409 and 2.121 ร
for high and low-spin, respectively). Slightly better results are obtained with OLYP/BS5, but the computational cost increases by a factor of 10. Thus, for the rest of this work we will use OLYP/BS4 as a method. Previously, we showed that in the [Mn(Cp n-Me ) 2 ] family (n = 0 to 5), it is possible to tune the ligand field via ring functionalization. Adding more methyl groups to the cyclopentadienyl ring increases the gap between the non-bonding and antibonding orbitals, switching from spin-crossover systems, such as [Mn(Cp) 2 ] or [Mn(Cp As can be seen from Table , in the low-spin state, the average value for d(Mn-Cp) remains almost unaltered, with an average value of (1.740 ๏ฑ 0.002) ร
(95% confidence). The same trend is observed for the high-spin state as well ((2.090 ๏ฑ 0.004) ร
, 95% confidence). However, although the geometry remains almost identical, large changes in the spin-state energy gap as well as the T 1/2 are observed. These changes can therefore only be attributed to the changes in the electronic structure of the [Mn(Cp 1-R ) 2 ] system introduced by the R group. |
60c75562842e654f73db4312 | 8 | In Figure , we plotted the computed T 1/2 against the ๏ณ p Hammett parameter for all the [Mn(Cp 1-R ) 2 ] systems in Table . As can be seen in the figure, a trend with the electron donating (or electron withdrawing) character of the R group can be observed. |
60c75562842e654f73db4312 | 9 | In general, a decrease in the T 1/2 is observed with increasingly withdrawing character of the R group. Although this trend in the computed T 1/2 spans on a 500 K range, the geometries of the computed systems remains almost identical, as can be observed using the average Mn-Cp centroid-distance. For all systems in Table , this parameter has average values of 1.734 ร
and 2.092 ร
for the low-and high-spin states, respectively (standard deviations of 0.0047 and 0.0082, see Supporting Information). This clearly shows the electronic nature of the tuning effect that the R group has on the SCO properties of the [Mn(Cp R ) 2 ] family. family (M = Cr II , Mn II , Fe II and Co II ). The computed energy gaps are in all cases larger than for the [Mn(Cp 1-Me ) 2 ] case (see Supporting Information), but the [Cr(Cp 1-Me ) 2 ] is close enough so one can envision a tuning of the ligand field to achieve SCO. Therefore, we replaced the methyl group by electron withdrawing groups (R=NO 2 and CN) |
60c75562842e654f73db4312 | 10 | Results for such calculations are summarized in Table . Although these substituents have a similar effect on the [Cr(Cp 1-R ) 2 ] as the one observed in [Mn(Cp 1-R ) 2 ], i. e. they reduce the spin-state energy gap, the corresponding T 1/2 does not experience a significant decrease. To understand that behaviour, we must analyze the corresponding entropy change in both families. While for the [Mn(Cp 1-Me ) 2 ] ๏S = 19.87 cal K -1 mol -1 , for the chromium analog [Cr(Cp 1-Me ) 2 ] this value is only ๏S = 11.61 cal K -1 mol -1 . This reduction of the entropy change is largely due to the electronic contribution to the entropy change, ๏S elec , that can be calculated |
60c75562842e654f73db4312 | 11 | . For Mn II , there is a large change in the total spin between high-spin and low-spin states (S HS = 5/2, S LS = 1/2), but for Cr II , the situation is much softer. Actually, the computed ๏S elec for Mn II and Cr II are, respectively, 5.96 and 3.31 cal K -1 mol -1 . Because T 1/2 = ๏H/๏S, for the chromium systems, the reduction of the spin-state energy gap is largely compensated by the decrease in the entropy change due to the Cr II electron configuration. Nevertheless, if we can reduce the ๏H even more for a [Cr(Cp 1-R ) 2 ] system, it should, potentially, exhibit SCO behaviour. |
60c75562842e654f73db4312 | 12 | Following this idea, we introduced a second NO 2 substituent, and computed the spinstate energy gap for the [Cr(Cp 1,3-NO 2 ) 2 ] molecule. For that molecule, ๏H=3.62 kcal mol -1 and ๏S = 10.0 cal K -1 mol -1 , thus making its T 1/2 = 362 K, a value that makes it a potential candidate for exhibiting SCO behaviour. |
60c75562842e654f73db4312 | 13 | Our working hypothesis is that, in principle, one can tune the ligand field around the metal center by making the ring more negatively charged, as is experimentally observed when moving from [Mn(Cp 1-Me ) 2 ], an SCO molecule, to the fully methylated system [Mn(Cp*) 2 ], a low-spin system. Following this idea, different substituents with more or less electron withdrawing effects have been tested in the [Mn(Cp 1-R ) 2 ] system. Although a trend is observed, the results are far from showing a clear correlation between the computed T 1/2 and the Hammett parameter ๏ณ p . However, a look at figure 1 allows us to identify several outliers that greatly deviate from the trend. |
60c75562842e654f73db4312 | 14 | We can thus see that the use of the Hammett parameters to describe the electron-donating character of the substituents and its effect on tuning T 1/2 holds to a great extent, but some points are clearly out of the trend. To understand the origin of such deviations, we collected more data on the low-spin state systems. In particular, NEVPT2 calculations analyzed using the ab initio ligand field theory (AILFT) framework were used to get the energies of the five d-based molecular orbitals, and a Hirshfeld charge analysis was done to calculate the total charge of the Cp 1-R rings for each R. Numerical data can be found in the SI. |
60c75562842e654f73db4312 | 15 | In this work, we presented a robust computational methodology to study the spincrossover behaviour of manganocenes of general formula [Mn(Cp 1-R ) 2 ]. The combination of the OLYP exchange correlation functional with a triple-๏บ basis set with polarization functions on all atoms provides a computed T 1/2 in excellent agreement with the available experimental data for the [Mn(Cp 1-Me ) 2 ] system. Also, geometrical parameters are properly reproduced. Single functionalization of the cyclopentadienyl ring allows for a fine-tuning of the SCO properties in this family. A general trend is observed with the electron-donating character of the R group. In general, the more electron-donating the R group is, the higher the T 1/2 . The electronic modulating effect of the R group is validated by the almost non-existing changes in the geometries of the studied systems, all of them having a very similar Mn-Cp distance in both spinstates. The use of R groups to increase or decrease the transition temperature opens the door for the use of other metals aside from manganese, which so far is the only one able to exhibit SCO in metallocenes. In particular, we showed that chromocene has a spin-state energy gap that can, in principle, be tuned using electron-withdrawing R groups such as CN or NO 2 . Although single functionalization of the Cp ring with such substituents does indeed reduce the spin-state energy gap, it is not enough to achieve a reasonable T 1/2 due to the much lower electronic entropic contribution from Cr II compared to the one provided by Mn II . However, this can be bypassed by introducing a second NO 2 group to the ring. Thus, our calculations not only allow for a rational understanding on how to tune the SCO properties of the [Mn(Cp 1-R ) 2 ] family, but also open the door for the computational design of new metallocenes with tailored spin-crossover properties that expand the current library of such compounds. |
61894fce92abe0f30f39d079 | 0 | bonds, showing remarkable levels of regioselectivity even in complex drug-like molecules (Scheme 1C). In HAT photocatalysis, a catalyst converts light energy into chemical energy for the homolytic cleavage of strong aliphatic C-H bonds. Especially, the decatungstate anion ([W10O32] 4-) has shown remarkable selectivity for specific C(sp )-H bonds, governed by an intricate balance between steric and electronic interactions. We envisioned that the regioselective introduction of an allyl moiety onto hydrocarbon frameworks would be particularly useful as it provides a convenient branching point for further late-stage synthetic exploitation (Scheme 1B). To install such moieties, radical allylation has manifested itself as a valuable strategy. One approach relies on the use of transition metal complexes to activate a substrate containing an allylic leaving group to afford a ๏ฐ-allyl complex, which is then suited to trap a C-centered radical (Scheme 1D). This strategy can engage a diverse set of allyl coupling partners but typically requires purposely designed radical precursors, which prevents the direct allylation of unactivated C(sp )-H bonds. SOMOphilic allylation constitutes another tactic and exploits radicofugal groups X (e.g., X = halide, SO2R, SnR3) in the allylic position to afford the desired product via a radical addition/fragmentation process (Scheme 1D). However, while synthetically useful, the reaction scope is mainly restricted to the synthesis of 1,1-disubstituted olefins. Seeking to address these challenges, we sought to develop a robust and versatile synthetic platform for the allylation of strong aliphatic C(sp )-H bonds. Hereto, a modular synthetic sequence is preferred in which the allyl moiety is assembled in a stepwise fashion, enabling the rapid generation of structurally diverse analogues. Specifically, our sequence features the merger of two distinct synthetic steps to accomplish this goal (Scheme 1 E). First, we planned to activate C(sp 3 )-H bonds via decatungstate-catalyzed hydrogen atom transfer and subsequently trap the resulting C-centered radical with a vinyl phosphonate. The ensuing radical addition product serves as a suitable linchpin for the second step, in which a classical Horner-Wadsworth-Emmons (HWE) olefination is able to deliver the targeted allylated compounds. In order to streamline these two steps, we reasoned that a telescoped flow protocol would be indispensable not only to accelerate access to these valuable building blocks but also to ensure facile scalability. Herein, we report the successful realization of such a flow platform enabling both early-stage and late-stage allylation of a wide range of hydrocarbons. During the writing of this manuscript, Silvi et al. reported a one-pot strategy to obtain 1,2-disubstituted olefins via a visible-light driven decarboxylative strategy merged with the Wittig reaction. Our investigations commenced with the decatungstate-enabled hydroalkylation of ethyl 2-(diethoxyphosphoryl)acrylate (2) using cyclohexane as the H-donor (See Supporting Information, Table ). Following a careful optimization of different reaction parameters, we found that the photocatalytic radical addition performed optimal in continuous-flow using a commercially available Vapourtec UV-150 photochemical reactor (PFA (perfluoroalkoxy) capillary, ID: 0.75 mm; V = 3.06 mL, flow rate = 0.612 mL min -1 , ๏ดr = 5 min) equipped with a 60 W UV-A LED light source, which matches the measured absorption spectrum of decatungstate. A 65% NMR yield (64% after isolation) was obtained for the targeted hydroalkylated compound when a CH3CN solution of the acrylate (0.1 M), cyclohexane (20 equivalents) and tetrabutylammonium decatungstate (TBADT, (Bu4N)4[W10O32]) as the photocatalyst (1 mol%) was irradiated for 5 minutes (See Supporting Information, Table , Entry 9). Other HAT photocatalysts, such as Eosin Y, anthraquinone, 5,7,12,14pentacenetetrone and fluorenone, were also evaluated, but failed to deliver the targeted product. Interestingly, benzophenone showed a comparable activity to the decatungstate anion, although only when used at high catalyst loading (20 mol%, 68% NMR yield). In addition, since benzophenone also dimerizes to give pinacol upon UV-A irradiation, we selected TBADT as the ideal photocatalyst for the targeted hydroalkylation reaction. Notably, this transformation is quite general and a diverse set of alkylphosphonates (3) could be readily isolated and characterized (see Supporting Information, Section 6). |
61894fce92abe0f30f39d079 | 1 | Next, the obtained alkylphosphonates were subjected to the successive HWE olefination (Scheme 2). A telescoped flow approach was developed in which the two individual steps were connected in a single streamlined flow process without intermediate purification. We selected 1,3-benzodioxole (1a), a common moiety in many medicinally-relevant molecules, as the Hdonor and exposed it to the photocatalytic reaction conditions. Upon exiting the photochemical reactor, the reaction mixture containing the alkylphosphonate is merged with a stream containing paraformaldehyde (3 equiv.) and lithium tert-butoxide (1.1 equiv.) in tetrahydrofuran. The combined reaction mixture is subsequently introduced into a second capillary microreactor (PFA, ID: 0.75 mm; V = 7.1 mL; ๏ดr = 5 min) and, after only 5 minutes of residence time, the targeted C(sp 3 )-H allylated product 4 could be obtained in 80% overall NMR yield (70% isolated yield). Notably, the tactical combination of these two steps in flow results in a very efficient and operationally simple protocol, delivering these coveted scaffolds in only 10 minutes overall reaction time. As another benefit, the flow process could be readily scaled to produce 5-10 mmol of the desired compound (65% isolated yield, Scheme 2) without the need for tedious reoptimization of the reaction conditions, which is typically associated with batch-type scale up procedures. |
61894fce92abe0f30f39d079 | 2 | To further demonstrate the potential of this operationally facile approach to introduce allyl functional groups, we wondered whether paraformaldehyde-d2 could be used in the HWE step. Such a straightforward, regioselective introduction of deuterium atoms in organic molecules would be of tremendous importance for mechanistic, spectroscopic and tracer studies. Using our two-step flow protocol, the analogous deutero-allylated compound 4-d 2 was isolated in 68% yield, perfectly matching the results obtained for the non-deuterated version 4. |
61894fce92abe0f30f39d079 | 3 | Similarly, N-Boc piperidinone was a competent substrate for this protocol affording the deuterated product 20 in 44% yield. Finally, in an effort to demonstrate the applicability of this method to the late-stage functionalization of some medicinally relevant molecules, we subjected several biologically active molecules to our two-step flow protocol. N-methyl-2pyrrolidone, often used in the formulation of drugs for both oral and transdermal delivery routes, could be regioselectively functionalized at the endocyclic ๏ก-to-N position (21, 52%). Also the terpenoid ambroxide (22, 40% yield) and the nootropic drug aniracetam (23, 20% yield) could be efficiently decorated with a deuterated allyl moiety. In a similar vein, we turned our attention to introduce aromatic and aliphatic aldehydes in the second step, yielding trisubstituted allyl moieties, which are particularly challenging to synthesize. By exploiting our modular protocol, a virtually limitless array of substituents can be systematically introduced (Scheme 3). Due to steric hindrance, prolonged reaction times (~ 3 hours) were required to obtain full conversion and thus a fed-batch approach was adopted (Scheme 2). Using this strategy, the expected olefin 24 was obtained in 60% yield (d.r. 2:1) when the reaction stream exiting the photoreactor was added to a stirring solution of benzaldehyde (1.5 equiv) and LiOtBu (1.1 equiv) in tetrahydrofuran. In general, aromatic aldehydes bearing electron-withdrawing substituents required shorter reaction times (e.g., 26-30) and the presence of ortho-substituents resulted in higher E-to-Z ratios (e.g., 31, 33 and 36). |
61894fce92abe0f30f39d079 | 4 | This allowed us to utilize our telescoped flow strategy (as shown in Scheme 2) for electronpoor aldehydes and, to our delight, similar results were obtained as with the fed-batch procedure (see e.g., 28, 29, 35-37). Notably, different classes of hydrogen donors, such as hydrocarbons (39, 43%), (thio)ethers (40-41, 47-68%), protected amines (42, 51%) and amides (43, 55%), proved all competent reaction partners. In all cases, the reaction performed particularly well, delivering densely functionalized alkenes in high yields and Estereoselectively. It is important to note that it would be extremely challenging to access either of these with the current radical allylation methodologies, which do not allow to synthesize trisubstituted alkenes (Scheme 1D). Unfortunately, all attempts to install fully-substituted olefins, by engaging ketones in the HWE step, failed. Interestingly, our protocol was also amenable to aliphatic aldehydes containing enolizable positions (44-48, 57-71% yield). The use of protected piperidine-4-carboxaldehyde allowed to obtain the corresponding allylated products 47 and 48 in excellent yields (60-68 %) and with good diastereomeric ratios. As a testament to the power of this strategy to rapidly diversify double bonds, medicinal agents and natural products containing carbonyls, such as acetyl-protected helicin, citronellal and indomethacin aldehyde derivatives, were also reactive delivering the targeted value-added olefins in synthetically useful yields (49-51, 20-63%). The regioselective and late-stage installation of allyl groups opens up innumerable possibilities for further diversification. As an illustration of this synthetic potential, we explored diverse conditions for the conversion of synthon 4 into functionalized derivatives (Scheme 4). The olefin and the ester functionalities could be orthogonally reduced by exploiting different reduction conditions, yielding compounds 52 (70%) and 53 (62%), respectively. Moreover, compound 4 was an ideal substrate for another Giese-type radical addition using decatungstatephotocatalyzed HAT (54, 62%). Finally, product 55 could be obtained via a classical Mizoroki-Heck-type coupling (60%). 58 In conclusion, we have developed a practical methodology which enables the modular and regioselective allylation of C(sp 3 )-H bonds. Our strategy involves a synergistic merger of a photocatalytic Hydrogen Atom Transfer and an ensuing Horner-Wadsworth-Emmons olefination in a scalable and telescoped flow protocol. In its present form, the synthetic platform offers rapid access to various di-and tri-substituted olefins from abundantly available hydrocarbon feedstocks, including biologically active molecules. The operational simplicity of our flow protocol, requiring no intermediate purification, should facilitate a rapid transition from academic to industrial settings. We anticipate that this practical method will unlock new synthetic opportunities for the rapid and late-stage diversification of building blocks, medicines, natural products and other specialty chemicals. |
60d8940e67d4915cb1962c4b | 0 | In recent years, crystal engineering and construction of coordination networks with fascinating structural topologies have attracted great attention owing to their potential as functional materials. Concurrently, the development of multidimensional networks based primarily on linking metal centers with rigid bridging components, such as 4,4`bipyridine has been initiated. Since then chemiluminescence's of 3-aminophthalhydrazide was first investigated by Albrecht, the acyl hydrazides are growing interest in the development of luminescent materials due to their potential application in emissive devices. Moreover, the attractive and promising practical applications in many other areas have stimulated further investigations in light-emitting devices, nonlinear optics, and functional films, conjugated polymers, logic functions of molecularscale, uranyl salts, saccharides, and aromatic organic molecular crystals which exhibit tribofluorescence or tribophosphorescence from the molecules comprising the crystal and/or nitrogen emission triboluminescence, azopolymers. Metal directed assembly have been used to generate luminescent materials based on complexation of transition metal and multifunctional bridging ligands and is one of many useful strategies to design extended frameworks of various topology and dimensionality. Ligands with amino group of anthranilic acid derivative backbones may affect the properties of the resulting emissions by promoting the coupling of metal atoms through its emission systems. It has been of interest to use photoactive ligands as building blocks to generate supramolecular polymers. Some of these workers have been concerned with the influence of structural changes upon the chemiluminescent properties. |
60d8940e67d4915cb1962c4b | 1 | The luminescence is appreciably enhanced when the hydroxyl or amine groups are in the ortho position. The efficiency of chromogenic sensing is more remarkably affected by the chemical environment of the anthranilic acid derivatives, depending on the presence of a protecting free amino group or amide group. Taking into account these considerations, it is possible to optimize the uv-vis sensing of the molecules by structure modification. Keeping the above applications in mind, we have investigated the synthesis and characterization of a series of 2,6-and 2,3-diamino benzoic methyl esters derivatives with various electron-donor groups at the 2,6-or 2,3-positions. |
60d8940e67d4915cb1962c4b | 2 | 3-nitropthalic acid and its anhydrides was chosen as a starting material due to its ready availability and to the fact that it can be easily converted to diamine derivative 5 and 10, a compound that can be selectively functionalized at the amino positions. The electron donor groups chosen for this study were pyridyl, 1,3-dithiolanyl, benzyloximate, and thiophenyl ether, all known for their Lewis basicity and ligating potential. This investigation reflects our ongoing interest in synthesizing new organic ligands for supramolecular design. The initial step is the functionalization described herein was the conversion of 3-nitropthalic anhydride and 3-nitropthalic acid to 2,6-diamino benzoic acid methyl ester (5) and 2,3-diamino benzoic acid methyl ester (10) 3-nitrophthalic anhydride under methanol reflux conditions gives mixture of 1 and 6 in ratio of 9:1 (by NMR) in 98% yield. The mixture on treatment with thionyl chloride under heating condition gives corresponding mixture of acid chlorides 2 and 7 from which the acyl azide 3 and 8 were prepared by treating with sodium azide in acetone in 90%. The mixture of acyl azides 3 and 8 under Curtius rearrangement followed by column chromatography purification gives two isolated product 4 and 9 in 70% and 5% respectively. The compound 4 under palladium charcoal treatment gives product 5 in 90% yields. The compound 9 was independently prepared in four steps by regioselective esterfication of 3nitro phthalic acid, followed by its acid chloride 7 in 95% yield, which on acyl azide conversions 8 in 70% yield, followed by and Curties rearrangement. In these it is 100% regioselective conversion, no trace of its corresponding isomer was detected even in crude NMR. The compound 9 under reduction by using palladium charcoal condition gives the 2,3-diamino benzoic acid methyl ester 10 in 80% yield, and acetonide protected diamine 16 (in 5%), is obtained as byproduct. The mechanism is unclear. |
60d8940e67d4915cb1962c4b | 3 | Derivatives 13-14 were each synthesized in straightforward two step procedures from 5 and 10 as summarized in Scheme 1. The diamines 5 and 10 on treatment with ๏ก-bromo acetyl chloride to give the bromo derivatives 11 and 12 in 69% and 66% yields, respectively. The bromo derivatives 11 under treatment with 2-pyridyl methanol gives the pyridyl derivative 13 as an oil in 59% yield after chromatographic purification. Attempts to prepare 14 by reaction of 12 with 2-pyridyl methanol in the presence of triethylamine or NaH were unsuccessful and resulted in the isolation of pyridyl ester product with elimination of aromatic moiety (by nmr), it is understood that due to the steric hindrance, elimination of aromatic moiety might occur. (Scheme 3). Reaction of 5 and 10 with 2.2 eq of 1,3-dithiolane-2-carbonyl chloride in the presence of triethylamine in THF led to the isolation of the bis(1,3-dithiolanyl) products 15 and 16 as white solids in 55% and 73% yield after purifications using column chromatography. The 16 structure is also confirmed by single crystal data. (Scheme 4). In a similar procedure, compound 5 was reacted with 2.2 equivalents of ๏ก-benzyloximino acid chloride . After chromatographic separation, 2,6-bis-benzyloxy methylester 19 was isolated in 66% yield along with the 6-mono-benzyloxy methylester 20 in 5% yields. Both compounds were isolated as white solids. (Scheme 6). We also attempted to synthesize the 2,3-bis-benzyloxy methyl ester 22 from direct nucleophilic substitution reactions of diamine 10 with ๏ก-benzyloximino acid chloride in either of TEA or pyridine and sodium hydride method. In both cases, our attempts were unsuccessful and resulted in either the isolation of starting materials or unidentified mixtures of products. But we were successful in synthesizing the 22 in two steps via acetonide protection (Scheme 7). Thus, reaction of 10 with 2,2-dimethoxy propane in presence catalytic amount of para toluene sulfonic acid gave 21 as residual oil in 47% yield after column chromatography purification, and reaction with ๏ก-benzyloximino acid chloride and workup with dil HCl gave 22 as white solid in 78% yield after chromatographic purification (Scheme 7). It is worth to mentioned that the while doing the Curtius rearrangement on compound 3, on heating at 110 o C gives urea derivatives, which is common intermediate step for the synthesis of quinazoline derivatives which having biological properties . Under acidic condition 24 gives benzo[d] [1,3]oxazin-4-one and in basic condition gives 2,4-quinazolindione derivatives respectively (scheme-9). |
60d8940e67d4915cb1962c4b | 4 | Preliminary experiments in which these compounds were mixed with various silver salts resulted in the formation of insoluble suspensions that could not be recrystallized for X-ray crystallographic analysis. However, elemental analysis of the isolated suspensions confirms the presence of silver, suggesting that complexation reactions are indeed occurring. The optical properties of compounds are of primary concern in chromogenic sensing and patterning. The solution-phase UV-vis absorption spectrum was recorded at room temperature in dilute chloroform solutions of about 10 -5 M concentration. As shown in table, In the absorption spectra for all compounds shows around 192 nm band seems to be an inherent property of these compounds and is not due to impurities as was verified by thin layer chromatography and NMR spectroscopy. The absorption maximum of 5 and 10 was remarkably blue shift shows at 348 nm each. But 2,6, disubstituted derivatives 15 and 19 shows absorption maximum at 271 and 332 nm respectively, the decrease of absorption maximum to 270 in case of 15 can be explained that it contains dithialone moiety, is responsible for it. Similarly, the compound 20 with a free amine group exhibited max at 360 nm, which is close to its parents compound 5 which having two free amines. Compound 19 compared with the absorption of corresponding its isomer compound 23 |
60d8940e67d4915cb1962c4b | 5 | We have prepared a series of new compounds containing dithiolane or benzyloxy oximate to aromatic rings bonded to substituted anthranilic acid core. The synthetic procedure is straightforward, and the products are obtained in good to excellent yields after chromatographic purification. Excellent solubility properties and the presence of electron donor groups in the modified anthranilic acid derivative described herein may be advantageous in applications such as anthranilic acid derivatives-metal conjugate synthesis, for example. The efficiency of chromogenic sensing is more remarkably affected by the chemical environment of the anthranilic acid derivatives, depending on the presence of a protecting amino group or amide group. Taking into account these considerations, it is possible to optimize the uv-vis sensing of the molecules by structure modification. All these compounds possess multiple sulfur atoms and are thus capable of binding in a multidentate fashion to soft transition metal ions. A reaction of these ligands with late transition metal ions is a current focus in our laboratory. |
60d8940e67d4915cb1962c4b | 6 | This procedure is modified from an earlier reported procedure. In the same manner described for 2 and 6, 2carbomethox-3-nitro benzoic acid (6) (4.5 g, 20.0 mmol) and thionyl chloride The reaction mixture was stirred at room temperature for 24 h. The THF was removed using a rotary evaporator and the residue was quenched by the addition of water (100 mL). The aqueous solution was extracted with ethyl acetate (100 mL), and the organic layer was washed with water (100 mL), saturated sodium bicarbonate solution (100 mL), water (100 mL), and brine (100 mL), and then dried over anhydrous MgSO4. After filtration and concentration of the filtrate, the residue was purified by column chromatography (20:80 EtOAc:hex) to give 13 as residual oil (0.67 g, 59%). Triethylamine (1.67 mL, 1.21 g, 12.0 mmol) was added and the solution was stirred for 15 minutes. 1,3-dithiolane-2-carbonyl chloride (1.22 g, 7.2 mmol) was dissolved in dichloromethane (5 mL) and the solution was added dropwise over a period of 15 minutes. The reaction mixture was stirred in the ice bath for 1h and stirred at room temperature overnight, and reaction mixture was quenched with aqueous solution of HCl and extracted with dichloromethane (50 mL X 2), and the organic layer was washed with water (100 mL), saturated sodium bicarbonate solution (100 mL), water (100 mL), brine (100 mL), dried over anhydrous MgSO4 and filtered. All volatiles were removed and the residue was purified using column chromatography (20: Triethylamine (365 mg, 3.6 mmol) was added, and the solution was stirred for 15 minutes. ๏ก-benzyloximino acid chloride (650 mg, 3.2 mmol) was dissolved in dichloromethane (5 mL) |
60d8940e67d4915cb1962c4b | 7 | and the solution was added dropwise over a period of 15 minutes. The reaction mixture was stirred in the ice bath for 1 h and stirred at room temperature overnight, and reaction mixture was quenched with aqueous solution of HCl and extracted with dichloromethane (50 mL X 2), and the organic layer was washed with water (100 mL), saturated sodium bicarbonate solution (100 mL), water (100 mL), brine (100 mL), dried over anhydrous MgSO4 and filtered. All volatiles were removed and the residue was purified using column chromatography (15:85 |
62a3951999b7836e25fbe614 | 0 | Molecular dynamics (MD) simulations have been used as a powerful computational tool by researchers in the fields of physics, chemistry, biology, and materials engineering, to model dynamical behavior of several material systems, including gases, liquids, solids, surfaces, and clusters. The accuracy of all these simulations and subsequent calculations is determined by the quality of the potential energy surface (PES) description, in the form of a force field model used in the simulation. The force field model describes the inter particle interactions in the system in question, and it is the most computationally expensive aspect of the MD simulation. High fidelity techniques, for instance, density functional theory (DFT) and ab-initio molecular dynamics (AIMD) are computationally expensive and are limited to much shorter time (โผ ps) and lengthscales (โผ nm). Hence the developement of an accurate classical description of the PES would be beneficial. Empirical potentials such as the Embedded atom model (EAM) which are derived from concepts based on the density functional theory, and its variants like Suttona) Also at Center for Nanoscale Materials, Argonne National Laboratory, Argonne, IL 60439 Chen (SC) and Gupta potentials, , are computationally efficient and have been successful in predicting much of the bulk properties including the elastic coefficients, thermal expansion coefficients , melting points , free energies in both solid and liquid phases , description of defects and many more for metals and metal alloys. |
62a3951999b7836e25fbe614 | 1 | Bond-order based empirical force field models (EFF) (e.g, Tersoff-type BOP, Reactive force field model (ReaxFF) ) are yet another class of interatomic potentials that attempt to account for bond directionality in materials via an angular dependence. The existing set of parameters for these pre-defined models perform well for bulk structures. But in recent years with the advancements in synthesis and applications of nano system materials, the need for having a computational route to assist with the design of novel low-dimensional materials has become increasingly crucial. But these popular and widely available force-field models which have been successful for bulk systems fail to capture the properties of these clusters accurately (e.g. see Fig 1 shows comparison of errors in prediction of energies and forces between some best stateof-the-art potential models vs. DFT energies for a wide size range of cluster configurations sampled from near to far-from equilibrium). Most of the elemental systems have mean absolute errors (MAE) around ยป 0.5 eV/atom illustrating the need for better models that can describe the interactions in lowdimensional systems. (meV/ร
) using force field models available in literature such as EAM/fs , EAM/alloy , MEAM , ADP , SC/EAM ) |
62a3951999b7836e25fbe614 | 2 | The need to accurately describe the potential energy surface of low dimensional systems including sub-nanometer clusters of transition metals with small number of atoms is critical given their exotic optical electronic magnetic as well as catalytic properties that are diverse and in many instances dependent on the cluster size. Their structural and dynamical properties are known to significantly different from their larger nanosized or bulk counterparts. This is attributed to the significantly higher surface to volume ratio, enhanced presence of dangling bonds, quantum confinement effect, as well as other electronic effects associated with the presence of d-orbitals which gives these clusters some unique chemistry . With greater advances in computational resources and advancements in ab initio calculations, high throughput computational approaches to materials design have been increasingly implemented in recent times. More recently, the advances in machine learning and data science have provided a new pathway to computationally explore the dynamical behavior as well as characterize the structure and energetics of these materials and accelerate the discovery of new materials for the various applications . A complete understanding of the energetic pathway, for example, would give us greater control over processes during an application and allow us to prevent undesirable side reactions . |
62a3951999b7836e25fbe614 | 3 | Beyond energetics, the dynamic evolution of these nanoclusters is another important factor from a materials design perspective. In several applications (e.g. catalysis or energy storage), these clusters often undergo structural and dynamical phase transitions from a stable configuration to another. One of the popular approaches to probe the dynamics of these clusters is Ab-Initio Molecular Dynamics (AIMD) . But despite the advancements in computational resources both in terms of hardware and software, the AIMD simulations are limited in the timescales and length-scales that they can access. This also limits the use of high-fidelity calculations such as DFT, in the use of exhaustive structure searches even with the most efficient sampling methods (e.g., evolutionary algo-rithms, basin-hopping, etc.). However, in classical molecular dynamics the atomistic and molecular interactions is such nanoscale systems continues to be described by models that were originally trained to bulk properties. |
62a3951999b7836e25fbe614 | 4 | A description based on classical mechanics, of the potential energy surface of the nano and sub-nano clusters of these metals provides us with a computationally cheaper alternative pathway for long timescale calculations of the cluster dynamics or carry out an exhaustive search of the structure/compositional space of these transition metal nano clusters. They face challenges because they trade in the accuracy of calculations for computational efficiency. The pre defined functional forms of the popular force field models used in classical molecular dynamics routinely fail in capturing the structure dynamics of the sub nanometer clusters with 10 to 100 atoms in its structure, even though they provide excellent results for the bulk properties of the same elements . The use of predefined functional form imposes serious limitations on the physics and chemistry that can be captured. This is because of the complexity of the potential energy surface (PES) can only be captured accurately by a very extensive set of basis functions and the popular empirical force-field models often use only a sub-set of this complete basis set. This causes the current physical models to be very accurate in some region on the PES but fail in others since the model does not account for the basis functions that describe the physics of that region. This becomes increasingly evident when we go to the nanoscale potential energy surface where the higher surface to volume ratio and the presence of dangling bonds and enhanced complexity in structural arrangements. Modeling such low-dimensional systems calls for a more diverse and complete basis set in the classical models so that the physical properties can be captured accurately. |
62a3951999b7836e25fbe614 | 5 | Neural network (NN) based potential models offer an alternative that is flexible enough to describe accurately the size dependent structural and dynamical properties in these nano and sub-nanoscale metal clusters. More recently, NN based potential models are emerging as a popular alternative to the classical functional based models, as a result of the rapid developments happening in the computational resources in terms of better and faster hardware as well as the vast selections of electronic structure codes that facilitate us in the efficient generation of the structural training dataset. The primary goal of these training schemes is to train effective NN potentials to achieve the accuracy and precision of high-fidelity ab-initio methods at a fraction of their computational cost. An inherent limitation of these NN models is that they are interpolative. The traditional approach to resolve this problem has been a brute force one where the NN is trained against as large a training dataset as possible. Such large-scale generation of high-fidelity training data can become challenging depending on the level of the electronic structure calculations employed. To tackle the challenges faced in the generation of an optimised training dataset, there have been several recent efforts to develop active learning and reinforcement learning strategies that facilitate in the effective sampling and selection of the training structures for NN models. It is an effort to ensure that the trained NN describes the potential energy surface for the elements accurately with minimum training dataset. For example, an active learning (AL) strategy based on the Query by Committee (QBC) scheme 51 was developed by Smith et al. which uses the disagreement between an ensemble of ML potentials to probe the reliability of the model's prediction. This approach allows for automatic sampling of the regions of chemical space which was not being described accurately by the ML potential model. It was validated over a test dataset of a diverse set of organic molecules and their results showed that one requires only 10% to 25% of the data to accurately represent the chemical space of these molecules. In another approach an AL scheme (deep potential generator (DP-GEN)) that constructs ML models for simulating materials at the molecular scale was introduced by Zhang et al. Their procedure involve exploration, generation of accurate reference data, and training. They used Al and Al-Mg as two representative cases and showed that accurate ML models can be trained with minimum number of reference data. In a recent work, unsupervised machine learning (ML) scheme coupled with a Bayesian optimization technique that evaluates the Gaussian Approximation Potential (GAP) model, which used dataset generated from a "meltquench" ab initio molecular dynamics (AIMD) was demonstrated by Sivaraman et al. From the trajectories of both melt and quench cycles of the AIMD simulations, the underlying cluster structures were exploited by partitioning them to "N" uncorrelated clusters. This modified dataset was sequentially sampled to train the GAP model until the desired accuracy was achieved. In another approach Vandermause et al. uses structures sampled on-the-fly from AIMD and an adaptive Bayesian inference method to automate the training of low-dimensional multiple element interatomic force field models. Their AL framework uses inherent uncertainty of a Gaussian process regression model to decide acceptance of model prediction or the need to augment training data. In all of the above studies, the overarching aim in these studies is to minimize the ab-initio training data required to describe the potential energy surface. |
62a3951999b7836e25fbe614 | 6 | Here, we introduce a new workflow that combines transfer learning (TL) with active learning (AL) strategy 55-57 -we first learn the PES from existing high-quality physics-based models in literature which describes the bulk systems efficiently, and subsequently augment this learning via retraining with a higher fidelity albeit sparsely sampled first-principles training dataset to concurrently capture both the nanoscale and bulk PES. Our workflow departs from status-quo in its ability to transfer the learning from decades of high-quality bulk interatomic potentials and subsequently improve via learning from a sparsely sampled dataset that nonetheless covers a diverse range of cluster configurations from near-equilibrium to highly non-equilibrium. Our workflow deploys transfer and active learning strategies that iteratively improves the fingerprinting depending on model fidelity. It allows us to train NN that learns the description of bulk and nanoscale potential energy surface from minimal amount of first-principles training data which is sampled on-the-fly from Nested Ensemble Monte Carlo simulations. Briefly, we introduce a two prong strategy (i) active learn the pair-wise interactions from best spherically symmetric potential models (ii) improve the learning by introducing 3-body and 4-body interactions against an additional first-principles sampled dataset. In active learning level, we start with the a minimum amount of training data (โผ 1 to 3 data points) which are continuously updated using data points sampled with a Nested Ensemble Monte Carlo scheme that iteratively queries energy landscape and retrains the network with an updated training set that includes failed configurations/energies from previous iteration until convergence is attained. The final training set at the time of convergence is used as starting dataset in the next level of training. |
62a3951999b7836e25fbe614 | 7 | To test the efficacy of our approach, we choose to model energetics of sub nanometer clusters of those elements which are traditionally described by the Sutton Chen (SC) type EAM potentials (Ag, Al, Au, Cu, Ir, Ni, Pb, Pd, Pt, Rh), and deploy the AL strategy to develop a NN potential model that can accurately predict the bulk as well as diverse geometries of nano clusters for those ten elements. For small cluster where bond-directionality effects are important, we show that popular spherically symmetric potential models such as embedded atom method (EAM) or its variants like Sutton-Chen (SC) and Gupta potentials, struggle to capture the diverse geometries and the dynamical properties. This is because of the presence of only a radially symmetric component in their functional form. Bond-order based empirical force field models (EFF) (e.g, Tersoff-type BOP, Reactive force field model ReaxFF ) account for bond directionality via an angular dependence. The existing parameterizations of these pre-defined models perform well for bulk structures, and close-packed (bulk-like) cluster configurations (at large sizes). For instance, the SC Potential works really well in predicting the bulk elastic properties of the ten FCC metals it describes. But they tend to fail when it comes to describing the potential energy surface for smaller sized clusters [ Fig 1]. These ten FCC metals represent excellent model systems to test our developed framework. We show that our TL and AL-NN are able to adequately represent the bulk and cluster energy landscape by sampling minimal amount of reference data (โผ 500-700 total structures in reference dataset). The performance of these NN in capturing bulk and cluster properties is discussed in detail. |
62a3951999b7836e25fbe614 | 8 | We introduce a workflow that trains high-dimensional NN to model bulk and cluster properties in 2 stages. In first transfer learning stage, we use the popular Sutton Chen type forcefields to calculate the energy of the configurations that are added to the training set. We first ensure that the Neural network trained using these structures would make predictions closely resembling the ones made by the SC EAM potential. An active learning approach is used to train the NN via 5 major steps: ( In stage 2, the final training set sampled from stage 1 is taken and the energies of the configurations are recalculated using DFT. This training set is augmented with additional actively (on-the-fly) sampled DFT data (clusters vs energetics) until convergence to DFT predicted bulk and cluster energetics is obtained. In this stage, the NN employs a more flexible symmetry functions i.e. accounts for radial as well as angular symmetry functions. This process allows us to take advantage of the radial nature of the SC EAM potential (pair-wise) and in the second stage add corrections (3-body and 4-body) that account for the angular dependence for the potential energies for a nano cluster. The NNs constructed in this study were interfaced with the LAMMPS, VASP and Atomic Energy Network (AENet) software package 58 , which were adapted to implement the transfer and AL schemes outlined above. Simulations using these networks were carried out using AENet interfaced with the Classy Monte Carlo simulation software 59 to perform the sampling during the transfer and AL iterations. |
62a3951999b7836e25fbe614 | 9 | The initial training set was limited to a single atom configuration and a highly stable configuration, representing a reasonably low-energy configuration of the PES, to kickstart the training process. Although any number of seed structures can be used, we chose the least number possible to introduce minimal bias in our AL workflow. |
62a3951999b7836e25fbe614 | 10 | We use a NN architecture with 4 layers of neurons. The hidden layers made of 15 nodes each. For stage 1 where we train the networks to a training set obtained using SC EAM forcefields, we use 8 nodes in the input layer. For stage 2 where the network is trained to DFT PBE energies an input layer of 30 nodes is used. These input nodes hold the symmetry functions that represent the potential energy surfaces of the different elements under consideration. The output layer with one node represent the potential energy of a given configuration. Besides, the input layer and the hidden layers contain a bias node that provides a constant signal to all the nodes of its next layer. The spatial coordinates of an atom in a nano structural configuration are mapped into the rotational and translational invariant co-ordinates as |
62a3951999b7836e25fbe614 | 11 | For stage 1, we used 8 radial symmetry functions G 1 each with different ฮท values. This decision to use only the radial symmetry functions at this stage was taken to leverage the radially symmetric nature of the Sutton Chen EAM forcefields. For stage 2 along with the above radial symmetry functions 22 angular symmetry functions were also used with a distinct set of parameters. This allowed us to map the increased complexities of the potential energy surface described by DFT for each of the elements in this study. |
62a3951999b7836e25fbe614 | 12 | For this study we represent each atom of a structural configuration by a neural network with the identical architecture and weight parameters. The total potential energy of the of that structure is defined as the sum of the values given by the output nodes of each of the NN. In every NN, all the compute nodes in the hidden layers receive the weighted signals from all the nodes of its previous layer and feeds them forward to all the nodes of the next layer via an activation function as |
62a3951999b7836e25fbe614 | 13 | A Levenberg-Marquardt approach 62 was used to optimize the neural network weights for each AL generation. The initial training set was limited to a single atom configuration and a highly stable configuration. For each of the active learning cycles the network was trained for 2000 epochs each representing a complete training cycle. The training scheme in AENet implements a k-fold cross validation scheme, where only a fraction of structures in the training set are used for objective minimization. The structures not used for the objective minimisation are used to cross validate the network to prevent over fitting, and the network with the least error with the cross validation scheme was chosen to be best network of the cycle and carried forward. |
62a3951999b7836e25fbe614 | 14 | Following the selection of the best network, a series of simulations are run to actively sample the configurational space described by the selected NN as shown in Fig. . For the purposes of this work in the first stage, a Boltzmann based Metropolis Monte Carlo sampling and a modified nested ensemble based approach 63 were used to generate the structures for each AL iteration. It was found in past research ( 64 ) that simple thermal sampling is not sufficient in properly exploring the configurational space. The success of an active learning scheme is based on the algorithm's ability to probe for configurations that are incorrectly predicted. Thermal sampling tends to sample only over a narrow range of thermally relevant structures. However, naive sampling approaches such as purely random cluster generation with no care given to energetics is often no better as this encounters the opposite problem in that it over biases the sampling toward entropically favored configurations while ignoring lower energy and other physical relevant states. As such the Modified Nested Ensemble which works by continuously halving the density of states was found to be a nice compromise between pure thermal or purely random sampling. The Metropolis simulation was run for 5, 000 MC cycles with 30 moves per cycle at 300K with the initial structure being randomly picked from the current neural network training pool. The Nested Ensemble simulations were run for another 24,000 cycles with 30 moves per cycle. Such a larger sampling through the nested sampling was required such that for elements with configurations with more diverse energy ranges would be covered completely and the structures representing the large range may be tested against the NN in the next step F. Testing the NN After the stochastic sampling step is completed, a set of 10 structures are gathered from the trajectory files of the Metropolis and a set of 15 structures are gathered from Nested Sampling trajectory files. These are sampled by dumping a structure every 1000 cycles for the metropolis run and for the nested sampling run dumping a structure every 1600 cycles. For the nested sampling run this is set up such that we pull one structure from each energy "strata" as the nested sampling gradually constricts the energy space. This ensures we are always testing structures ranging from very high energy down towards more equilibrium and low energy regions of the configurational phase space. The primary goal of this is to expose structures that the model incorrectly labels. In stage 1 the real energies of the structures were calculated using Sutton Chen EAM potentials and compared with the predictions of the NN model. Each time the energies do not agree within a specific tolerance that structure is added to the training set of structures. here we use a variable tolerance scheme where the tolerance of each cycle is determined as the 105% training error for that cycle the value for which can be obtained from the AENet log files for the training. This ensures that the structures whose energies are predicted catastrophically wrong compared to the training error are added into the training set each cycle. This ensures that in the initial stages of active learning all relevant structures in both the highly stable regime as well as the higher energy configurations get added to the training set at a higher pace as well as, in the later stages of the active learning slows down the addition of structures to the training set, reducing the risk of the network overfitting. This entire process is continued until the exit criteria is hit. For this work, we specified that if no new structures were added in 9 consecutive AL iterations, that the potential has converged. The converged NN potential following the stage 1 was validated extensively with a validation set of 35000 structures which spanned the entire range of energies the network was trained in. |
62a3951999b7836e25fbe614 | 15 | The reference energies and forces of all the cluster configurations were computed using density functional theory (DFT) as implemented in the Vienna Ab initio Simulation Package (VASP) . The low energy cluster configurations are further relaxed using conjugate gradient approach 66 whereas in the case of non-equilibrium cluster we perform static DFT calculation to compute their energies and forces. To avoid interactions between two periodic images we assign the box length such that the distance between two periodic images is at least 15 ร
or larger. The DFT calculations were performed using the generalized gradient approximation (GGA) with the Perdew-Burke-Ernzerhof (PBE) exchange correlation functional, and the projected-augmented wave (PAW) pseudo-potentials . The details of pseudo-potentials used in our calculations are provided in the Table in Supplementary Information Section 1. Spin polarized calculations with the Brillioun zone sampled only at the ฮ-point were performed. To handle errors that may arise during the structural relaxation or static DFT calculations, our high throughput workflow used an in-house python wrapper around VASP along with a robust set of error handling tools. |
62a3951999b7836e25fbe614 | 16 | For stage 2, we perform transfer learning by taking the converged training set from stage 1 and recalculating their energies using DFT-PBE. The NN was re-trained using revised fingerprint files which accounted for the radial as well as the angular symmetry operations. The training set was augmented with a set of 100 highly structurally optimized clusters as well as the structures representing the EOS for each of the elements. This was implemented in order to further facilitate the NN in accurately predicting the forces for highly stable structures as well as the bulk elastic constants. The re trained networks were also validated extensively with validation sets of 2000 -3000 structures. To evaluate the equation of state (EOS) plot for all these FCC metals, ยฑ 5% strain was applied in all three directions. The initial bulk structure of fcc system for each of these elements has been collected from materials project database. Three independent elastic constants, were determined for cubic system by employing suitable lattice distortions represented by a strain tensor, ฮต in such a way that the new lattice vectors r' in the distorted lattice is given by r' = (I + ฮต)r where I is the unit matrix. This method was adapted from our previous work and the workflow was modified to perform both the active and transfer learning stages. |
62a3951999b7836e25fbe614 | 17 | We evaluate the performance of the first stage of our Transfer/Active learning workflow, where the NN was trained to the SC EAM potential. As part of the training, note that only the radial symmetry functions were taken into consideration using the fingerprint file. We can also observe that the error drops rapidly in the initial stages of the AL and going forward, the addition of more structures slows down and subsequently, the fine tuning of the network begins. Eventually the NN reaches a stage where more structures are no longer added i.e. the NN errors are within the specified tolerance (< 20 meV/atom). The size of the training set as well as the number of AL iterations needed to reach the convergence for each of these elements are influenced by a variety of factors including the range of energies over which the configurations are sampled (from near to far-from-equilibrium) and the complexity of the configuration space the network is trying to learn. |
62a3951999b7836e25fbe614 | 18 | Fig. represents a correlation plot that compares the performance of the final optimized NN on the SC training set. The AL-NN predictions of energies and the forces for the clusters in the training pool are compared with the reference SC EAM model predictions. We find that the MAE during training is ยซ 20 meV/atom for all the systems and those during testing are also ยซ 20 meV/atom for most elements except Ir and Rh, which have marginally higher error owing to increasing configurational diversity in the far-from-equilibrium region. The predicted MAE in forces are much lower than 0.5 eV/ร
for all elements. We also further test the NN using a validation set of 35000 structures never used in the training process. Fig. (b) and (d) show the correlation between AL-NN predicted energies and forces, respectively vs. reference SC EAM model. As expected, we find that the final optimized NN is able to reliably predict the cluster energies for an elaborate test data set generated not only near equilibrium, but also in the highly non-equilibrium region that extends far beyond. As a more rigorous test of the performance of AL-NN, we compute the forces on the atoms for the various clusters in the validation set and compare those with that obtained from SC EAM. It should be noted that the forces were not included as part of the training during the AL iterations. Fig. 4 (d) shows the correlation between the AL-NN predicted vs. SC EAM forces. Each point in this correlation plot represents one of the force components -F x , F y and F z acting on a particle. We find the overall MAE between AL-NL vs SC EAM predicted forces is relatively small i.e MAE ยซ 1 eV/ร
for most elements with Ir and Rh displaying somewhat higher errors (consistent with the trend in the MAE for energies). Given that the NN had not been trained on the forces, this agreement is of excellent quality. |
62a3951999b7836e25fbe614 | 19 | We next compare the SC trained NN and compare its performance with energetics and forces computed with DFT. While the SC model is expected to deviate from the DFT energies for certain situations, it is important to quantify the extent of error of the SC trained NN against a DFT test set. To accomplish this, we further test these networks with the validation set of 2000 -3000 structures whose energies were calculated with DFT to probe into how well a network trained with only radial symmetric functions will perform with respect to DFT. We find that the NN fails catastrophically in both energy as well as force calculations (Fig. ) -the MAE for most elements are ยป 0.5 eV. The corresponding MAE in force predictions shown in Fig. (b) are also quite high for most elements (ยป 1 eV/ร
). This provides hints to the inadequacies of an SC EAM type of potential (spherically symmetric), which accounts for only the radial symmetry operations, in describing the energetic as well as dynamics of systems. At the nanoscale, as surface effects begin to dominate, it becomes imperative to include contributions from angular i.e. 3-body interactions as well. |
62a3951999b7836e25fbe614 | 20 | Having established that the training with SC EAM as a reference is inadequate to accurately describe the nanoclusters, we proceed to transfer the learning based on SC dataset to that from a higher fidelity DFT dataset. We took the training set of the SC converged NN from the previous stage, recalculated their energies using DFT, augmented the training set with 100 highly stable bulk configurations and structures representing the equations of state (EOS) and retrained the network taking into account the angular symmetry functions. Fig. plots the correlation between the performance of the retrained NN on the new training set. The transfer learned NN predictions of energies and the forces for the clusters in the training pool are compared with the reference DFT model predictions. MAE for the training set was found to be less than 20 meV/atom, which is of the same order of magnitude as DFT error. |
62a3951999b7836e25fbe614 | 21 | We test the transfer learnt NN with a validation set of 2000-3000 structures that was not part of the training process. Fig. 6 (b) shows the correlation between AL-NN predicted energies vs. reference DFT energies. As expected, we find that the re trained NN is able to reliably predict the nano cluster energies. We observe significant improvement in the accuracy of energy as well as force prediction by the NN with respect to the DFT predictions as seen in Fig. ). |
62a3951999b7836e25fbe614 | 22 | Apart from describing the clusters accurately, it is also essential for the NN to appropriately describe the bulk properties. Fig 7 (a) shows the comparison between the lattice constants predicted by the NN with the ones predicted by DFT. We find that there is an excellent agreement between the NN predictions when compared to DFT. The Cohesive energies predicted by our NN is also in good agreement with the values predicted by DFT as shown in Fig . The correlation between the three unique elastic constants for the ten elements with an fcc crystal structure considered are given in Fig ). The elastic constants for bulk elements also show good agreement with the DFT predicted values even though we do not use elastic constants as part of our NN training set. The elastic constant C 44 posed the biggest challenge as accurately describing the value required the errors in the energies be far more smaller than what we had with our SC trained NN. To circumvent this issue, we augmented our training set with the structures representing the EOS of the elements. Table provides a summary of the performance of the neural network potentials on bulk mechanical properties. We find that the elastic constants are predicted within 10-15 % of the DFT value and were thus able to get an excellent agreement on all three elastic constants as well as the EOS between the ones predicted by NN vs. DFT (Fig. ). |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.