Inverse design framework using deep convolutional generative adversarial networks for light-trapping structures in perovskite solar cells

Thin-film solar cells require effective light-trapping structures to compensate for reduced optical path length in sub-wavelength active layers. This work presents a Deep Convolutional Generative Adversarial Network (DCGAN) framework for designing such structures, with physical validation performed using the Transfer Matrix Method (TMM). Optical constants for methylammonium lead iodide (MAPbI3) perovskite, including the full complex refractive index n(λ) + ik(λ) from published spectroscopic ellipsometry measurements, serve as the material basis. TMM computations for a flat 300 nm perovskite film on SiO2 yield 33.2% average absorption across 400–1000 nm. Addition of a 100 nm MgF2 anti-reflection coating raises performance to 39.4%, while an effective medium textured layer achieves 38.2%. Energy conservation (R+T + A=1) holds at all 31 wavelength points. The DCGAN architecture, trained on TMM-generated structural distributions, produces candidate geometries at 0.10 s per design following a one-time training phase of 1.5 h of CPU training. A performance-driven selection loop evaluates generated structures through TMM, retains high-performing candidates, and refines the generator iteratively. Spectral analysis confirms strong absorption at wavelengths where k exceeds zero (63% at 400 nm, 58% at 600 nm) with expected decline near the bandgap at 800 nm (5.4%). The trained DCGAN generates structures averaging 50.8% absorption, with the best design reaching 57.5%, representing a 24.3 %age point improvement over the flat baseline of 33.2%. This approach provides a material-agnostic platform applicable to emerging thin-film photovoltaics including perovskites and CIGS.