Machine learning accelerated nonadiabatic molecular dynamics of defect-mediated recombination in alkali metal passivated Cu2ZnSnS4

Abstract

Nonradiative electron–hole recombination mediated by defects is a critical loss mechanism in semiconductors and requires computationally demanding nonadiabatic (NA) molecular dynamics simulations. In this work, we develop a machine learning-accelerated framework to model long-time NA couplings in Cu2ZnSnS4 containing CuZn antisites and CuZn + ZnCu antisite pairs, with and without alkali metal passivation. By formulating NA coupling evolution as a time-series prediction problem, we benchmark 37 representative deep learning architectures, including recurrent neural networks, convolutional neural networks, transformers, and hybrid models. Among them, the extended long short-term memory model achieves the best overall performance, yielding an average test set R2 of 0.98 while maintaining high computational efficiency. This approach enables accurate reconstruction of long-time NA coupling trajectories at a prediction cost reduced by over five orders of magnitude relative to repeated direct first-principles NA coupling evaluations. The application of the framework shows that alkali metal doping systematically reduces the NA coupling strength and accelerates decoherence, with Li exhibiting the strongest suppression of recombination, which is associated with enhanced charge localization and lattice fluctuations. The predicted recombination lifetimes agree closely with first-principles results, validating the reliability of the machine learning approach. This work establishes a generalizable strategy for machine learning-assisted NA dynamics simulations and provides mechanistic insights into defect-mediated recombination processes.