PNNL

Abstract

A physics-regularized machine learning (ML) approach is developed for predicting time temperature-transformation (TTT) diagrams from alloy composition in uranium-based systems. To ensure physically realistic C-curve morphology while maintaining predictive accuracy, four loss function formulations: (1) baseline mean squared error (MSE), (2) MSE regularized with a semi-empirical nucleation-based model, (3) MSE regularized with shape constraints that enforce characteristic C-curve convexity, and (4) a hybrid approach that combines both semi-empirical regularization and shape constraints are systematically evaluated. Models incorporating shape constrained loss functions produce TTT predictions with physically consistent C-curve topology and quantitative agreement with experimental isothermal transformation data. Post hoc model explanation using global feature importance and partial dependency plots reveals systematic changes in the learned composition-TTT relationships as a function of the loss function formulation. Applying the optimized model to previously unexplored U-Mo-X ternary systems identifies U-Mo-Pt, U-Mo-Au, U-Mo-Ta, U-Mo-W, and U-Mo-Fe as candidate alloys that extend the a-phase nucleation time relative to U-Mo binary baseline alloys. This finding indicates improved kinetic stability against ?- to a-phase decomposition during isothermal holds. Furthermore, nose time predictions from the hybrid physics-regularized model are compared with those from a Gaussian Process Regression (GPR) model using MSE loss. This physics-regularized ML framework offers potential for accelerating alloy discovery in composition spaces where traditional TTT experimental determination is time and resource-intensive.