Comparison among machine learning models for prediction performance of Seebeck coefficient

Abstract

The advancement of materials science has been propelled by the integration of computational methods such as molecular dynamics (MD) and density functional theory (DFT). Recently, machine learning (ML) has emerged as a promising tool, offering faster and accurate predictions by leveraging data-driven approaches. This study focuses on applying ML techniques to predict the thermoelectric properties of materials, particularly the Seebeck coefficient, a critical parameter for evaluating ther- moelectric efficiency. Utilizing the Ricci ab initio computational database, this research compares the performance of different ML models, including CGCNN and CraTENet, and explores a novel feature fusion model that integrates structural and compositional information. The results demonstrate that CGCNN outperforms CraTENet in predicting the Seebeck coefficient, benefiting from structural information. However, the proposed fusion model, while robust, exhibits intermediate performance between the individual models, suggesting that simple feature addition may not fully leverage the strengths of both architectures. Additionally, multi-task learning approaches for predicting both Seebeck coefficient and bandgap show slight performance trade-offs but highlight the potential of simultaneous property prediction. These findings underscore the importance of leverag- ing domain-specific correlations and advanced feature fusion techniques to optimize thermoelectric material discovery and design.