Phase diagram construction and prediction method based on machine learning algorithms

Phase diagram, which is known as the “compass” and “map” of materials research, plays a guiding role in the material design and development. Conventional CALPHAD method could provide the detailed information on the phase equilibria via the assessment of thermodynamics model parameters. However, CALPHAD assessment for the multi-component systems can be particularly challenging due to the significant time costs involved and lack of experimental data, especially when attempting to predict the phase diagrams of multi-component systems. The fast-growing machine learning technique opens a new pathway to deal with tons of data and parameters. Meanwhile, the CALPHAD method has accumulated abundant high-quality phase diagram data who would be the perfect training data for the machine learning algorithms. In the present work, a phase diagram prediction method which integrates machine learning algorithms with CALPHAD descriptors is proposed. The present study establishes and train machine learning models to predict phase-type and solvus temperature of the materials. Using training datasets obtained from the CALPHAD method, we combine the total Gibbs energy and magnetic descriptor with training set to predict the isothermal sections of Cu–Co–Ni and Fe–Cu–Co ternary systems. The results indicate that the elevated temperatures not only enhance the solubility of Co, Ni, and Cu in intermetallic compounds but also facilitate the formation of eutectic precipitates (γFe+αCo). This methodology can efficiently predict the phase diagram of material system with higher number of components by training the phase diagram data of lower ones, thereby providing a new strategy to complement the CALPHAD with machine learning technique and extend the application of CALPHAD method to the advanced materials including high entropy alloys and functional materials.