Machine Learning-Driven Prediction and Optimization of Thermoelectric Transport Properties in Ce-Doped CaMnO3 Ceramics

Abstract

Oxide-based thermoelectric materials are promising candidates for high-temperature waste heat recovery owing to their excellent thermal and chemical stability. Nevertheless, optimizing their thermoelectric performance remains challenging because electrical resistivity and the Seebeck coefficient exhibit complex, nonlinear, and composition-dependent relationships. In this work, experimental measurements combined with data-driven machine learning approaches were employed to predict and optimize the high-temperature thermoelectric behavior of Ce-doped CaMnO3 (Ca1–xCexMnO3, x = 0.02–0.10) ceramics using Support Vector Machine (SVM), Multilayer Perceptron (MLP), eXtreme Gradient Boosting (XGBoost), and a hybrid Convolutional Neural Network–Long Short-term Memory (CNN–LSTM) architecture. The models were trained and validated using experimentally measured electrical resistivity (ρ) and Seebeck coefficient (S) data collected over the temperature range of 50–800 °C. Among the investigated approaches, the hybrid CNN–LSTM model, which uniquely combines spatial feature extraction with sequential learning to resolve complex nonlinear transport behaviors that conventional machine-learning models fail to capture accurately, exhibited the highest predictive performance, achieving coefficients of determination (R2) of 0.9982 for electrical resistivity and 0.9987 for the Seebeck coefficient. Permutation feature importance analysis further revealed that the Ce concentration has a stronger influence on the transport properties than the operating temperature. To identify the optimum operating conditions, 150,000 Monte Carlo simulations were performed using the trained CNN–LSTM model to generate continuous multidimensional response surfaces. The optimization results identified a favorable thermoelectric region at a Ce content of x = 0.05–0.06 and temperatures between 350 and 500 °C, where the simulated power factor exceeded 0.35 mW/K2m. These results demonstrate the potential of hybrid deep learning for data-driven analysis and optimization of oxide-based thermoelectric materials under the investigated experimental conditions.