Bio-Inspired Crystal Machine Learning for Band Gap Prediction in Hybrid Halide Perovskite Monolayers

Abstract Bio-inspired sparse connectivity offers a compact route for exchanging global information in crystal learning, but it has rarely been explored for inorganic–organic solids. We develop FlyCrysNet for band gap prediction in 357 relaxed A2BX4 hybrid halide perovskite monolayers and evaluate it alongside 17 established crystal machine learning models. The 18 models represent descriptor-based neural networks, message-passing graph neural networks, crystal graph Transformers, physics-informed long-range and reciprocal-space models, and bio-inspired sparse graph networks. Their test root-mean-square errors (RMSEs) range from 0.222 to 0.405 eV. ReciNet gives the smallest error, with a mean absolute error (MAE) of 0.171 eV, an RMSE of 0.222 eV, and a coefficient of determination (R2) of 0.966. FlyCrysNet combines a periodic graph encoder with a fixed sparse feature mixer and learnable topology control, reaching a test RMSE of 0.227 eV and an R2 of 0.964. It ranks second by test RMSE and forms a high-accuracy group with ReciNet, PotNet, and iComFormer. Topology controls further show that the sparse mixing framework remains stable under self-only, random, degree-preserving, and unweighted variants. Two low-gap Sn–I monolayers identified from the learned chemical space were examined using the Heyd–Scuseria–Ernzerhof screened hybrid functional (HSE06). (iPA)2SnI4 and (EA)2SnI4 retain direct gaps at Γ points of 1.61 and 1.40 eV, respectively, with band edges dominated by the Sn–I framework. The results establish bio-inspired sparse feature mixing as an effective crystal learning strategy and connect accurate band gap prediction with focused electronic structure analysis of hybrid perovskite monolayers.