(611f) End-to-End Learning for Prediction of Optoelectronic Properties of Organic Photovoltaic Polymers

Despite algorithmic improvements to quantum-mechanical methods, computational design of atomistic systems is limited by the vast combinatorial space of possible molecular structures. Algorithms that efficiently interpolate more detailed simulations across this combinatorial landscape are therefore needed. Traditional approaches to building quantitative structure-property relationship models cast molecules to a fixed-size numeric vector using static methods that do not change during the learning process. In this talk, we demonstrate how end-to-end learning approaches [1] drastically improve prediction accuracy over fingerprint-based approaches for large datasets. We apply the method to a computational database for active-layer materials for organic photovoltaic solar cells. The database contains approximately 30,000 unique chemical compounds, for which optoelectronic properties were calculated using time-dependent DFT with the B3LYP/6-31g(d) model chemistry. Our method is able to predict HOMO, LUMO, and band gap values extrapolated to the polymer limit [2] using only the monomer’s structural information. Prediction accuracy is shown to have a root mean squared error of approximately 0.05 eV for HOMO, LUMO, and band-gap calculations while taking many orders of magnitude less time to perform. Additionally, we demonstrate the ability of the algorithm to learn on subsets of the data, and the amount of data required to fully parameterize end-to-end learning algorithms such that they outperform traditional fingerprinting-based methods.

[1] Gilmer J, Schoenholz SS, Riley PF, Vinyals O, Dahl GE. Neural Message Passing for Quantum Chemistry. arXiv. 2017;cs.LG. (2017).

[2] Larsen RE, J. Phys. Chem. C, 120, 9650-9660 (2016).