Deep learning framework for performance prediction and inverse design of nanowire solar cells

We propose a deep learning framework for nanowire array solar cell design that integrates forward prediction and inverse design. A hybrid convolutional neural network-transformer model is developed to predict photovoltaic metrics and full current-voltage characteristics with high accuracy, achieving a fairly low mean absolute percentage error of less than 1% (with a minimum of 0.04%) and a mean absolute error of 0.21 mA/cm², respectively. For inverse design, a conditional generative adversarial network is developed to handle the one-to-many mapping between performance and structure, and a candidate selection strategy is introduced to enhance design reliability. The results demonstrate strong agreement among the target performance, forward-predicted performance, and physics-based simulation results for the inversely designed structures. This work may pave the way for efficient and reliable data-driven performance prediction and inverse design of optoelectronic devices.