Machine Learning-Assisted Novel Photovoltaic Optimization for Tailored Ultra-Thin CdTe-Based Solar Cells

Abstract

This study presents a machine learning-based design framework utilizing deep Q-learning (DQL) to optimize ultra-thin CdTe solar cells with active layer thicknesses ranging from 100 to 400 nm. By coupling the transfer matrix method for optical analysis with SCAPS-1D simulations for electrical modeling, the DQL agent effectively explored the complex parameter space, optimizing the thicknesses of all key layers, including SnO₂, CdS, CdTe, MoO₃, and Au. The DQL framework intelligently adjusted each layer based on electromagnetic wave propagation and absorption profiles, enhancing internal reflection and light trapping within sub-micron geometries. Even at extremely low absorber thicknesses (e.g., 100 nm), the optimized structures achieved high photovoltaic performance, with power conversion efficiencies up to 9.39% and J_sc values exceeding 11 mA/cm². At 400 nm, efficiency increased to 15.75% with J_sc of 20.86 mA/cm². These results demonstrate that efficient photon harvesting and carrier transport are achievable through full-stack optimization. External quantum efficiency and absorption spectra confirmed the integrated optical-electrical enhancement achieved by DQL. This work highlights the capabilities of reinforcement learning in high-dimensional solar cell design problems and provides a scalable approach for developing next-generation, lightweight, efficient, and material-conscious photovoltaic technologies.