AI-Driven prediction system for Thermoelectric Properties: Unveiling Novel Materials for Sustainable Energy Solution in Nigeria

Abstract

Nigeria's persistent energy deficit continues to underscore the need for sustainable alternative energy technologies. Thermoelectric materials convert heat directly into electrical energy and offer a promising pathway for addressing this challenge; however, the discovery of high-performance compounds remains experimentally demanding. This study presents an AI-driven machine learning framework and web-based prediction platform for estimating five thermoelectric properties: Seebeck coefficient (S), electrical conductivity (σ), thermal conductivity (κ), power factor (PF), and figure of merit (zT) directly from chemical formula and measurement temperature. Seven regression models were trained on 5,205 experimental records spanning 880 distinct chemical compositions. An 88-feature engineering framework incorporating compositional descriptors, temperature transformations, complexity metrics, and interaction terms was developed, with variance thresholding reducing the feature set to 61 variables. XGBoost achieved the best predictive performance for zT (R² = 0.796, RMSE = 0.154), Seebeck coefficient (R² = 0.787, RMSE = 96.8 μV/K), electrical conductivity (R² = 0.823), and thermal conductivity (R² = 0.797), while Gradient Boosting yielded the highest performance for power factor (R² = 0.863).Tree-based ensemble methods consistently outperformed linear models, confirming the non-linear nature of structure–property relationships in thermoelectric systems. Measurement temperature, heavy-element ratio, Shannon compositional entropy, and temperature–composition interaction terms emerged as the most influential predictors. The results demonstrate the effectiveness of compositional machine learning as a rapid and cost-effective screening tool for thermoelectric materials discovery. The validated framework has been deployed as an open-access web application supporting both single-compound prediction and batch screening, with automated identification of promising candidates for experimental validation.