Physics-guided multi-task learning for predicting thermophysical properties of Ag-In-Cd absorber alloys with extremely small data

Machine learning (ML) techniques have recently shown great promise in advancing nuclear materials research. However, the application of ML in this field is often limited by the scarcity of experimental data due to the high cost and complexity of irradiation experiments. In this study, we propose a physics-guided multi-task neural network (PGMTNN) framework to accurately predict the thermophysical properties of Ag-In-Cd control rod alloys under neutron irradiation, using extremely limited datasets. The PGMTNN model incorporates fundamental physical laws as soft constraints within a multi-task learning architecture, enabling the simultaneous prediction of density, specific heat capacity, thermal diffusivity, and thermal conductivity. This approach effectively captures the intrinsic correlations among different material properties and ensures physical consistency of the predictions. Compared with traditional ML methods and the transformer-based foundation model, the PGMTNN demonstrates superior performance, achieving an average coefficient of determination exceeding 0.92 on the testing set. These results underscore the potential of physics-guided, multi-task ML frameworks for accurate property prediction and reliable design of advanced absorber materials under small-data conditions.