De Novo Design of Molecules with Low Hole Reorganization Energy Based on a Quarter-Million Molecule DFT Screen

Abstract

Materials exhibiting higher mobilities than conventional organic semiconducting materials such as fullerenes and fused thiophenes are in high demand for applications in printed electronics. To discover new molecules in the heteroacene family that might show improved hole mobility, three de novo design methods were applied. Machine learning (ML) models were generated based on previously calculated hole reorganization energies of a quarter million examples of heteroacenes, where the energies were calculated by applying density functional theory (DFT) and a massive cloud computing environment. The three generative methods applied were (1) the continuous space method, where molecular structures are converted into continuous variables by applying the variational autoencoder/decoder technique; (2) the method based on reinforcement learning of SMILES strings (the REINVENT method); and (3) the junction tree variational autoencoder method that directly generates molecular graphs. Among the three methods, the second and third methods succeeded in obtaining chemical structures whose DFT-calculated hole reorganization energy was lower than the lowest energy in the training dataset. This suggests that an extrapolative materials design protocol can be developed by applying generative modeling to a quantitative structure−property relationship (QSPR) utility function.

Design of Organic Electronic Materials With a Goal-Directed Generative Model Powered by Deep Neural Networks and High-Throughput Molecular Simulations

“…Yet, the approach is often challenged by the low success rate in generating valid chemical structures, let alone ones with the desired properties (Lim et al, 2018). On the other hand, we confirmed that over 80% of the two Optimization in the property space has also proven more successful with the RNN based generative design, as reported in benchmark studies (Brown et al, 2019;Marques et al, 2021).…”

“…In fact, we observe in Figure 6 the sign of extrapolation in the glass transition temperature with the newly generated compounds, as the validation set analyzed in this work has 17 compounds with T g greater than the upper limit from the training set (215°C). This is supported by the previous study where the REINVENT formalism was used to create the new molecular designs ( Marques et al, 2021 ), reporting the model’s general capability of extrapolating outside the training set domain. Based on our observation, we believe the best strategy to bring the extrapolation capability to the generative model is 1) to take advantage of multiparameter optimization, 2) to maximize the number of training set data to provide enough resolution in the property space to extend near the edges of the training set domain, and 3) to have the property prediction model—such as quantum chemistry for electronic properties and QSPR for glass transition used in this work—available well outside the training set space to keep the accuracy of predicted properties for the validation sets high.…”

“…On the other hand, we confirmed that over 80% of the two million-plus SMILES strings generated by the prior network in this work were valid chemical structures, and rendering the rate of developing valid chemistry a non-issue for this framework. Optimization in the property space has also proven more successful with the RNN based generative design, as reported in benchmark studies ( Brown et al, 2019 ; Marques et al, 2021 ).…”

De NovoDesign of Molecules with Low Hole Reorganization Energy Based on a Quarter-Million Molecule DFT Screen: Part 2

“…To perform HRE predictions, we precisely followed our previously reported approach, , repeated here for reference, with minor changes.…”

“…The HRE in the training datasets were in the range of 0.0734–0.2958 eV for the 500 point set, 0.0558–0.3998 eV for the 50k set, and 0.0548–0.490 eV for the 250k set. The distributions of the sets are shown elsewhere . Each model was created using Schrödinger’s automated quantitative structure–activity relationship (QSAR) tool, AutoQSAR/DeepChem .…”