Machine learning assisted discovering of new M2X3-type thermoelectric materials

Recent years have witnessed a continuous discovering of new thermoelectric materials which has experienced a paradigm shift from try-and-error efforts to experience-based discovering and first-principles calculation. However, both the experiment and first-principles calculation deriving routes to determine a new compound are time and resources consuming. Here, we demonstrated a machine learning approach to discover new M2X3-type thermoelectric materials with only the composition information. According to the classic Bi2Te3 material, we constructed an M2X3-type thermoelectric material library with 720 compounds by using isoelectronic substitution, in which only 101 compounds have crystalline structure information in the Inorganic Crystal Structure Database (ICSD) and Materials Project (MP) database. A model based on the random forest (RF) algorithm plus Bayesian optimization was used to explore the underlying principles to determine the crystal structures from the known compounds. The physical properties of constituent elements (such as atomic mass, electronegativity, ionic radius) were used to define the feature of the compounds with a general formula 1M2M1X2X3X (1M + 2M: 1X + 2X + 3X = 2:3). The primary goal is to find new thermoelectric materials with the same rhombohedral structure as Bi2Te3 by machine learning. The final trained RF model showed a high accuracy of 91% on the prediction of rhombohedral compounds. Finally, we selected four important features to proceed with the polynomial fitting with the prediction results from the RF model and used the acquired polynomial function to make further discoveries outside the pre-defined material library.