The High-dimensional Phase Diagram and the Large CALPHAD Model

In the intricate world of alloy design, the CALPHAD (Calculation of Phase Diagrams) method has long served as a foundational pillar, shedding light on the complex phase statuses across varied compositions and temperatures for decades(1). Historically, alloy design has largely been reliant on binary or ternary phase diagrams(2-4). While these diagrams have been invaluable, they exhibit a distinct limitation: their inability to adequately address Complex Concentrated Alloys (CCAs), constrained by the representation of just two compositional variables per figure(5). This shortcoming necessitated the adaptation of algorithmic approaches for CCAs design(6-8). However, these algorithm-driven approaches, like the genetic algorithm, tend to provide a non-systematic design, often gravitating towards local optima and limited compositions that satisfy the criteria. The essence of their inefficiency boils down to a dual problem: the inherent limitations of computational speed and a point-by-point validation approach that inherently lacks systematic exploration. To truly revolutionize alloy design, one must address both the speed and the systemic design hurdles. The advent of GPU-accelerated machine learning offers a tantalizing solution for the speed conundrum(9, 10).