Physics-Guided Inverse Estimation of Standard-Specimen-Equivalent UTS from Subsized Tensile Measurements Using Explainable Machine Learning

Standard tensile testing is often impractical for irradiated nuclear materials and other high-value components because of limited material availability and specimen-extraction constraints. Subsized specimens provide a practical alternative, but their tensile response can differ from that of standard specimens because of geometry-dependent constraint, strain localization, and size effects. This study presents a physics-guided, explainable inverse machine-learning framework for estimating standard-specimen-equivalent ultimate tensile strength (UTS) from subsized tensile measurements using a Size Effect Ratio (SER), defined as the ratio of reference standard UTS to subsized UTS. A literature-derived dataset of 1050 observations and 55 variables was screened to 861 observations. Physics-guided descriptors representing grain-size strengthening, geometric constraint, plastic-zone effects, irradiation severity, strain-energy response, and solid-solution strengthening were incorporated into a CatBoost model, with shapley additive explanations (SHAP) used for interpretation. Because the YS/UTS descriptor contains the measured subsized UTS and therefore has an intrinsic relationship with the SER target, the target-independent formulation excluding YS/UTS was considered the primary performance assessment. Under leakage-controlled 10-fold cross-validation, this formulation yielded R2=0.386, MAE = 0.0945, and root mean square error (RMSE) = 0.1263. The corresponding formulation including YS/UTS achieved R2=0.849, MAE = 0.0368, and RMSE = 0.0626 and was treated as an auxiliary sensitivity analysis. Validation using 21 directly paired standard/subsized observations yielded R2=0.697, MAE = 0.0716, and RMSE = 0.0893 for the paired analysis. Genuine out-of-fold standard-UTS reconstruction achieved R2=0.485, MAE = 63.73 MPa, and RMSE = 88.53 MPa. These results indicate that SER-based inverse learning can provide an interpretable data-driven approach for estimating standard-specimen-equivalent UTS from subsized measurements, while also highlighting the limitations associated with constructed reference values and target-dependent descriptors.