Machine learning potential model for accelerating quantum chemistry‐driven property prediction and molecular design

Abstract Quantum chemistry (QC) calculations have significantly advanced the development of materials, drugs, and other molecular products. Molecular geometry optimization is an indispensable step for QC calculations. However, its computational cost increases dramatically with increasing molecular system complexity, hindering the large‐scale molecule screening. This work proposes a deep learning‐based molecular potential energy surface prediction tool (DeePEST) to significantly accelerate geometry optimizations. The key of DeePEST involves the development of a novel machine learning potential model for accurate and fast predictions of molecular energy and atomic forces. These predictions enable efficient molecular geometry optimizations for subsequent predictions of QC properties (single‐point energy, dipole moment, HOMO/LUMO, and 13 C chemical shifts) and COSMO‐SAC‐based thermodynamic properties (activity coefficient). Moreover, DeePEST facilitates efficient computer‐aided molecular designs that involve QC‐based geometry optimizations. The utilization of DeePEST in geometry optimizations achieves high prediction accuracy approaching to rigorous QC methods while maintaining the computational efficiency of molecular mechanics methods.