Efficient Configuration Sampling for Hybrid Functional DFT Calculations to Train Machine-Learning Potentials: Application to Atmospheric Chemistry

Abstract

Machine-learning potentials (MLPs), trained to predict energies from quantum chemical calculations, are widely employed to conduct large-scale MD simulations. However, MLPs are mostly trained on computationally inexpensive local/semilocal functionals, as generating training datasets using higher-accuracy theories, such as hybrid functionals, is challenging due to their high computational cost. Here, an active transfer learning scheme is developed to efficiently sample configurations for hybrid functional calculations. The proposed method is evaluated on atmospheric secondary aerosol formation reactions: clustering of sulfuric acid (SA), and dimethylamine (DMA), and oxidation of toluene. The accuracy of the trained MLP is shown to be comparable to that of the quantum chemical calculations, with errors within a few meV per atom. The molecular dynamics simulations are executed stably on a nanosecond scale, resulting in the formation of nanometer-size clusters. Thus, this study paves the way for establishing a general protocol to enable high-level atomistic simulations for a wide range of chemical systems.