Understanding the excited state properties of molecules provides insight into how they interact with light. These interactions can be exploited to design compounds for photochemical applications, including enhanced spectral conversion of light to increase the efficiency of photovoltaic cells. While chemical discovery is time- and resource-intensive experimentally, computational chemistry can be used to screen large-scale databases for molecules of interest in a procedure known as high-throughput virtual screening. The first step usually involves a high-speed but low-accuracy method to screen large numbers of molecules (potentially millions), so only the best candidates are evaluated with expensive methods. However, use of a coarse first-pass screening method can potentially result in high false positive or false negative rates. Therefore, this study uses machine learning to calibrate a high-throughput technique [eXtended Tight Binding based simplified Tamm-Dancoff approximation (xTB-sTDA)] against a higher accuracy one (time-dependent density functional theory). Testing the calibration model shows an approximately sixfold decrease in the error in-domain and an approximately threefold decrease in the out-of-domain. The resulting mean absolute error of ∼0.14 eV is in line with previous work in machine learning calibrations and out-performs previous work in linear calibration of xTB-sTDA. We then apply the calibration model to screen a 250k molecule database and map inaccuracies of xTB-sTDA in chemical space. We also show generalizability of the workflow by calibrating against a higher-level technique (CC2), yielding a similarly low error. Overall, this work demonstrates that machine learning can be used to develop a cost-effective and accurate method for large-scale excited state screening, enabling accelerated molecular discovery across a variety of disciplines.
Bibliographical noteFunding Information:
S.V. was supported by the Marshall Scholarship. This work made use of the CX1/2 high-performance computing clusters at Imperial College London.82 The authors would like to thank Daniel Davies for useful discussions surrounding machine learning, Kyle Swanson for help and advice about using Chemprop, Jiali Li for helpful discussions regarding active learning, and Martijn Zwijnenburg and Stefano Angioletti-Uberti for fruitful discussions about computational chemistry. The authors also acknowledge the UK Materials and Molecular Modeling Hub for computational resources, which is partially funded by Engineering and Physical Sciences Research Council (EPSRC) (Grant Nos. EP/P020194/1 and EP/T022213/1).
© 2022 Author(s).