Accelerating Energy Materials Discovery: AI-Driven Prediction of Thermoelectric Properties

The discovery of high-performance thermoelectric materials is crucial for sustainable energy applications, but the traditional trial-and-error approach is time-consuming and resource-intensive. This study presents a robust, machine-learning framework for the rapid computational screening and design of thermoelectric materials based on the dimensionless figure of merit (ZT). Using an ensemble method that strategically combines Random Forest, XGBoost, and Neural Network models, we enable relative screening of ZT trends from a set of 45 descriptors composition-based electronic, thermal, and structural descriptors. Trained on 115 experimentally measured ZT values, the ensemble achieved a modest predictive accuracy (test set R2 = 0.687, MAE = 0.162) and enabled effective relative screening across a broad chemical space. High-throughput evaluation of approximately 500 candidates successfully identified 20 promising materials, with op candidates such as SnSe_single and PbTe-Na identified with screening-level ZT indicators exceeding 2.0. Interpretable machine-learning analysis revealed two critical design principles: maximizing atomic mass variance to suppress lattice thermal conductivity and optimizing the electronegativity difference (Δχ) between 0.5 and 1.0 to enhance the electronic power factor. This study demonstrates how an integrated informatics approach can accelerate the discovery of novel thermoelectrics and provides a data-driven framework for guiding the rational design of high-performance energy materials.