Enhancing classical Scheil–Gulliver model calculations by predicting generated phases and corresponding compositions through machine learning techniques

Abstract

The classical Scheil-Gulliver model is an important tool for simulating non-equilibrium solidification processes in materials science, especially for rapid cooling processes such as additive manufacturing. However, the high computational intensity of the Scheil-Gulliver calculations through the CALculation of PHAse Diagrams (CALPHAD) method, especially for complex alloys, limits its application in high-throughput scenarios. This study introduces a novel machine learning (ML)-based approach to enhance the calculation of the Scheil-Gulliver model, facilitating efficient and large-scale simulations. We developed a suite of ML models to predict generated phases and their elemental composition in the Fe-Ni-Cr-Mn system. By integrating these models with a parallel calculation algorithm, the calculation process is completed in 52 minutes, while performing direct one-by-one calculations could take months. Our high-throughput calculations successfully processed 176,688 out of 176,851 compositions. Based on the calculated data, an algorithm was designed for linear gradient pathway planning. Thirty pathways from the BCC_B2 phase to the FCC_L12 phase were used for exemplification, with 28 pathways validated as feasible.