Machine Learning-Based Composition Design of Functionally Graded Alloys

Functionally graded materials (FGMs) effectively alleviate residual stress induced by physical property mismatch at dissimilar material interfaces through a graded transition in composition or structure. Among these, the matching of the coefficient of thermal expansion (CTE) is a core indicator for ensuring the service reliability of the joint. Traditional composition design relies on empirical trial-and-error, which makes it difficult to efficiently identify the optimal path in a high-dimensional composition space. This study proposes a data-driven, machine learning-assisted composition design method. Based on a high-precision dataset covering 15 elements and 747 CTE data points, six typical regression models were systematically evaluated. The results show that the random forest (RF) model achieves the best performance, with a coefficient of determination (R2) of 0.929 and a root mean square error (RMSE) of 0.658 on the test set. Using the SHapley Additive exPlanations (SHAP) method, the lattice constant (c), Young’s modulus (YM), and temperature (T) were identified as the key physical descriptors governing the thermal expansion behavior. Experimental validation shows that the CTE prediction deviation of the model for the high-performance Fe-based alloy Norem02 in the range of 20–300 °C is only 0.89%. Based on this framework, the composition of the 316L/Norem02 transition layer was successfully designed in this study. This effectively reduced the interfacial thermal expansion mismatch. Consequently, it provides a reliable theoretical basis for the rational design of dissimilar material interfaces under extreme service conditions.