A machine learning framework for solar cell design optimization and efficiency prediction

Abstract This study applies machine learning (ML) to the design of tin-based perovskite solar cells (PSCs) to reduce dependence on exhaustive device simulation. Ten regression algorithms were trained on 4428 SCAPS-1D simulations covering 27 device architectures, predicting efficiency (η), open-circuit voltage (V OC ), short-circuit current density (J SC ) and fill factor (FF) from the absorber, electron and hole transport layer materials, absorber thickness, and band-edge densities of states. CatBoost performed best, reaching R 2 of 0.999990 and RMSE of 0.0126% points for η on a random 80:20 split. Accuracies of this kind are common in this literature, so we examined what they measure. The dataset is a dense sweep of a deterministic solver, so random partitioning, including ten-fold cross-validation, leaves near-identical configurations on both sides of the split. Withholding whole architectures raised the error by one to two orders of magnitude. The models are therefore reliable for interpolation within characterised architectures, not for extrapolation to unseen ones. Training on random subsets showed that 480 simulations, 10.8% of the grid, recovered the correct architecture and the optimum in all ten repeats. The best configuration, FASnI 3 with SnO 2 and Cu 2 O at 660 nm, reaches 28.87% efficiency (V OC 1.086 V, J SC 30.01 mA cm⁻², FF 88.55%).