Abstract
Machine learning can accelerate thermoelectric materials discovery, yet most models remain “black boxes” lacking physical insight. We present an interpretable framework coupling Extra Trees regression with SHAP analysis to predict electrical conductivity (EC), Seebeck coefficient (SC), and thermal conductivity (TC) across 3879 multicomponent compositions. Physics-inspired feature engineering expands descriptors to 73–84 per target via polynomial lattice-distortion terms, cross-interactions, and compositional disorder statistics. Random-split test-set 𝑅2 values reach 0.961, 0.965, and 0.958 for EC, SC, and TC, respectively, surpassing a lightweight deep neural network. Composition-wise splitting reduces 𝑅2 to 0.822, 0.871, and 0.876, indicating limited generalisation to unseen compositions. SHAP analysis reveals property-specific associations: elemental descriptors dominate all targets, lattice-distortion terms influence EC, and heavy-element features affect TC. As a consistency check, a ZT-proxy computed from the independently predicted transport properties agrees reasonably well with experimental values on the test set (𝑅2 = 0.84); it is intended as a screening aid rather than a verified figure of merit. This interpretable framework supports screening and prioritisation backed by, not replacing, experimental validation.