Backward Mapping from Device Targets to Chemical Genomes for Interpretable Discovery of Phase-Stable Lead-Free Double Perovskites with DFT-Validated Design Rules

Abstract

Lead-free halide double perovskites are promising alternatives to Pb-based semiconductors, but their discovery is challenging because structural formability, thermodynamic stability, band-gap placement, optical-transition strength, dielectric screening, and carrier transport must all be satisfied within the vast A2BB'X6 space. We present a backward-mapping, genome-guided framework linking device-level targets to chemically interpretable descriptor families for Pb-free double-perovskite discovery. From 13,088 charge-balanced compositions, we apply a halide-aware workflow integrating geometric formability filtering, six-family chemical-genome descriptor encoding, evolutionary-optimized machine learning surrogates, SHAP-based interpretation, and DFT phenotype closure. Stability is modeled using Ehull-derived labels, while a band-gap surrogate predicts scalar-relativistic PBE Eg for target-driven selection. The funnel reduces the search space to seven DFT-validated candidates: K2BePdF6, K2MnCdCl6, Rb2TeCuBr6, Cs2SnGeBr6, Cs2GeSrBr6, Cs2NiBaI6, and Cs2AgInCl6, all verified for structural assignability, band-edge character, effective masses, dielectric response, optical absorption, conductivity, reflectivity, energy-loss spectra, and XRD fingerprints. Functional rules emerge from stability-function coupling rather than band-gap optimization alone, providing an interpretable inverse-design paradigm to accelerate Pb-free double-perovskite discovery.