Deep Learning-Based Generation of Lead-Free Organic-Inorganic Hybrid Halide Perovskite Materials Using Conditional Variational Autoencoders

Abstract

Organic-inorganic hybrid perovskites (OIHP) are promising materials for photovoltaic applications. This study proposes a computational candidate-generation framework for discovering new lead-free OIHP compositions, targeting band-gap energy as the primary property, while volume per atom, atomization energy, and density were used as supporting material descriptors. The framework integrates two coupled models: an ANN-based predictor and a CVAE-based generator. The ANN predictor achieved R2 scores of 89%, 93%, 91%, and 93% for band gap energy, volume per atom, atomization energy, and density, respectively. The CVAE generator successfully produced lead-free perovskite candidate compositions with band gap energies ranging from 1.2 eV to 3.6 eV. DFT validation of the generated candidates showed that 10 of 11 compositions (90.9%) had band-gap deviations below 10%, with a mean deviation of 6.40%. The smallest deviation was 2.70% for CH3NH3CaI3, whose predicted band gap was confirmed by DFT calculation.