Inverse Problem Analysis of Phase Fraction Prediction in Aluminum Alloys Using Differentiable Deep Learning Models

In recent years, there has been an increasing demand for the optimizationOptimization of alloy properties, driven by the growing complexity of end products and the need to reduce development costs. In general, Thermo-Calc based on the CALPHADCALPHAD method, which calculates the thermodynamic state of an alloy, is widely used for efficient alloy development. However, a challengeChallenges in alloy exploration using Thermo-Calc is the need for specialized computational skills and the significant computational effort required due to the extensive range of calculation conditions for numerous alloys. Consequently, we have developed a deep learningDeep learning model that rapidly and accurately predicts the temperatureTemperature-dependent changes in equilibrium phase fractions for 6000-series aluminum alloysAluminum alloys (AlAl–Mg–Si-based alloys), which are widely used in industry, using calculations from Thermo-Calc. We developed the architecture of the deep learningDeep learning model based on the TransformerTransformer, which is commonly used in natural language processing tasks. The model is capable of performing calculations more than 100 times faster than Thermo-Calc. Furthermore, by leveraging backpropagation of errors in the trained model, we developed a method to estimate the alloy composition for the phase fraction results calculated based on Thermo-Calc.