Accelerating nonadiabatic molecular dynamics simulations in CdSexTe1-x solar cells with recurrent neural networks

CdSexTe1-x is one of the most commercially successful absorber materials for thin-film solar cells, where Se doping is crucial for enhancing device efficiency. However, the microscopic control mechanisms are not yet clear. This study combines density functional theory-based electronic structure with nonadiabatic molecular dynamics to systematically analyze the nonradiative electron-hole recombination of CdTe and Se-doped systems. The results show that Se doping reduces the NA couplings and accelerates phonon-induced decoherence, thereby extending the nonradiative recombination lifetime. To accelerate NA coupling calculations, we introduced three types of recurrent neural network models (LSTM, GRU, and ConvLSTM) for extrapolation predictions. Among them, ConvLSTM demonstrated the best performance in terms of prediction accuracy and reproducing nonradiative behavior. Further combined with intrinsic defect calculations, it was shown that Se doping effectively shallows the transition energy levels of VCd and VTe and suppresses the formation of deep-level traps. This study reveals the multiscale mechanism of Se regulation on the nonradiative recombination pathways of CdTe-doped systems and provides an efficient machine learning framework for accelerating NA simulation of optoelectronic materials.