Discovering virtual antiperovskites as solid-state electrolytes through active learning

In surveying an extensive library of 18,133 hypothetical antiperovskites (X₃BA), we address the challenges posed by conventional experimental and computational screening methods. We introduce a novel computational approach, leveraging an active learning framework that synergistically integrates genetic algorithms with Bayesian optimization. This method efficiently discerns thermodynamically stable antiperovskites, narrowing down the vast initial set to 43 compounds characterized by minimal energy above the hull (E_h<150 meV/atom). Subsequent evaluations of their band structure (E_g), electrochemical stability window (V_w), and room-temperature lithium ion conductivity (σ_RT), alongside the construction of a 4-dimensional Pareto frontier for E_h, E_g, V_w, and σ_RT, refined this list to 22 promising candidates, seven of which exhibit outstanding room-temperature ionic conductivity (>4 mS cm⁻¹), marking them as potential candidates. Our methodology not only expedites the identification of superior antiperovskites but also establishes a groundbreaking paradigm for computational exploration in materials science.