Explore the full temperature-composition space of 20 quinary CCAs for FCC and BCC single-phases by an iterative machine learning + CALPHAD method

The search for desired complex concentrated alloys (CCAs) remains a daunting task because of the vast temperature/chemical composition space. While CALPHAD is a reliable technique, it requires intensive computations. In contrast, machine-learning (ML) methods can be fast and efficient but rely on a large and high-quality dataset. In this work, we combine these two techniques by implementing a reinforcement learning strategy to accelerate the exploration of CCAs. Starting from an initial small dataset from Thermo-Calc calculations with TCHEA3 database, the reinforcement learning is performed iteratively with the XGBoost ML training/testing and CALPHAD verification to progressively augment the dataset. This strategy allows for the identification of all single-phase FCC and BCC structures in the temperature-composition space of 20 Al-containing quinary alloy families formed by Al, Co, Cr, Cu, Fe, Mn, Ni and Ti, and achieves testing accuracies of above 97% and 92% on Thermo-Calc and on experimental data, respectively. The data analyses show that these 20 families exhibit a large disparity in their single-phase formation ability with AlCoCrFeNi and AlCrFeMnNi having the highest formation ability for FCC and BCC, respectively. Remarkably, this large disparity can be well explained by refined phase selection rules and structural inheritance from binary and ternary systems. Our extensive analysis also reveals the rarity of single-phase CCAs at room temperature. The method proposed and the findings revealed present new dimensions for the design and engineering of CCAs.