Materials Expert-Artificial Intelligence for materials discovery

Advances in materials databases create an opportunity to uncover descriptors that predict emergent properties, yet most studies rely on high-throughput ab initio calculations that can diverge from experiment. Experimentalists instead depend on intuition honed by hands-on work. We present “Materials Expert-Artificial Intelligence” (ME-AI), a machine-learning framework that translates this intuition into quantitative descriptors extracted from curated, measurement-based data. Using a set of 879 square-net compounds described using 12 experimental features, we train a Dirichlet-based Gaussian-process model with a chemistry-aware kernel. ME-AI reproduces established expert rules for spotting topological semimetals (TSMs) and reveals hypervalency as a decisive chemical lever in these systems. Remarkably, a model trained only on square-net TSM data correctly classifies topological insulators in rocksalt structures, demonstrating transferability. Complementing electronic-structure theory, our framework scales with growing databases, embeds expert knowledge, offers interpretable criteria, and guides targeted synthesis, accelerating materials discovery and rapid experimental validation across diverse chemical families. Material databases offer avenues for identifying predictive descriptors, yet often rely on data that diverges from experimental results. Here, machine learning was used to capture expert intuition into quantifiable descriptors, revealing hypervalency as a key predictor for topological semimetals.