Nonradiative recombination dynamics simulations of CuIn1−xGaxSe2 solar cells based on RNN and transformer models

Abstract

Nonradiative electron–hole recombination is a key mechanism of energy loss in optoelectronic materials, and one of its core regulation mechanisms is nonadiabatic (NA) coupling. However, traditional first-principles-based NA coupling calculations are often accompanied by high resource consumption, which restricts their application in large-scale material systems. To this end, this work takes CuIn_1−xGa_xSe₂ (CIGS) as a model system, selects two representative components with Ga/(Ga + In) (GGI) ratios of 0.5 and 0.25, uses Hammes–Schiffer–Tully (HST) and norm-preserving interpolation (NPI) methods to obtain four types of NA coupling datasets, and develops a variety of deep learning models for efficient prediction. Among them, the RNN_FCNPlus model achieved the highest average determination coefficient R² (0.960) in all tasks and the TransformerRNNPlus model performed most robustly in the transformer class (average R² of 0.944). Substituting the predicted NA coupling into the non-equilibrium dynamics simulation, the nonradiative recombination lifetime obtained is highly consistent with the direct calculation result. On further combining with the spatially localized mode (SLM) and pure decoherence function analysis, it is found that with the decrease of GGI, the phonon spectrum of the system is in the overall low-frequency region and the coherence time is extended from 40.3 fs to 55.7 fs, indicating that the decoherence process in the system is effectively suppressed. Although the band gap shrinkage and coherence enhancement may promote nonradiative transitions to a certain extent, the sharp decrease in NA coupling strength (from 0.502 to 0.376 meV for HST and from 1.610 to 1.022 meV for NPI) becomes the dominant factor, which extends the nonradiative lifetime from 7.66 ns to 10.83 ns and from 0.86 ns to 1.71 ns in the two methods, respectively. The above results not only reveal the microscopic physical mechanism of Ga/In ratio regulation of nonradiative behavior, but also verify the reliability and practicality of deep learning in accelerating NA dynamics simulation.