How would AI shape CALculation of PHAse Diagrams?

This 2-page paper discussed the connection between current data-driven machine learning scheme and CALPHAD. From my own perspective, CALPHAD should be considered as one kind of machine learning tasks for materials thermodynamics function and phase diagram. The challenges of CALPHAD is exactly the challenge of data-driven machine learning, few data, the multi source of data may lead to inconsistency and it’s also a multi-object optimization. Besides, it is also part of science: the results should follow thermodynamics laws, the fundamental principles. Unfortunately, it seems that this kind of consideration is never widely acceptable both from CALPHAD community itself and other people. Although it’s rejected and not appreciated by one of AI4Science workshop’s attention track this year due to the lack of quantitative illustration, I consider it’s still useful to share the thoughts here and would not try to publish it.