Inverse Prediction of the CALPHAD-Modeled Physical Properties of Superalloys Using Explainable Artificial Intelligence and Artificial Neural Networks

Abstract

The CALPHAD methodology models the physical, mechanical, and thermodynamic properties of materials based on specified alloy compositions using phase equilibrium calculations and thermodynamic databases. With the CALPHAD approach, millions of material-property data can be obtained for each alloy composition over various temperature ranges. However, finding an alloy with the desired properties often requires lengthy trial-and-error processes that involve manually adjusting the composition. In this study, the goal is to inverse this approach using artificial intelligence to predict alloy compositions that yield the desired properties. Accordingly, in the JMatPro software based on the CALPHAD methodology, the physical properties (density, thermal conductivity, linear expansion, Young's modulus, bulk modulus, shear modulus, and Poisson's ratio) of 250 different Ni-Cr-Fe-based superalloys in the temperature range of 540–920 °C were modeled. A dataset with 5000 rows was created from the generated data, of which 80% was used to train the artificial intelligence model, while 20% was reserved for validation and testing. Through analyses using Explainable Artificial Intelligence (XAI) and Artificial Neural Networks (ANN), alloy compositions with the desired physical properties at a given temperature were predicted with a high accuracy rate of 98.03%. In conclusion, beyond obtaining material properties from alloy compositions through the CALPHAD approach, artificial intelligence techniques make it possible to accurately inverse predict alloy compositions that yield specified physical properties at a particular temperature.