Artificial Intelligence for Learning Material Synthesis Processes of Thermoelectric Materials

Abstract

Various calculation-based and data-driven methods have been proposed to discover high-performance thermoelectric materials for sustainable energy resources. However, although several data-driven methods successfully discovered the chemical compositions of promising thermoelectric materials, the practical potential of the existing methods is still limited because there is a complex engineering problem between the discovered materials and real-world material synthesis. To tackle the engineering problem in material synthesis, we propose a multimodal graph-to-sequence model that predicts necessary synthesis operations and their engineering conditions from the chemical compositions of the precursor and desired product materials. For an experimental evaluation of the proposed method, we constructed a benchmark dataset containing precursor materials, product materials, and synthesis processes of 771 unique thermoelectric materials. The proposed method achieved the prediction accuracy greater than 0.85 in Jaccard similarity and F1-score in a task of predicting material synthesis processes on the benchmark dataset. Furthermore, the proposed method successfully generated material synthesis recipes described in the human language via large language models (LLMs). The collected thermoelectric dataset and the source code of the proposed method are publicly available at https://github.com/ngs00/spende.