Optuna-driven interpretable model for refractory multi-principal element alloys yield strength prediction with small dataset

Refractory multi-principal element alloys (RMPEAs) consist of high-melting-point elements, exhibiting excellent high-temperature mechanical properties and rendering them promising candidate materials for the aerospace field. However, strength-ductility trade-off, vast compositional space, and focus on compressive properties in most studies lead to insufficient high-quality tensile data, forming a vicious cycle that limits data-driven design. This study addresses the critical challenge of accurate yield strength prediction for RMPEAs under small-dataset constraints while overcoming the opacity of black-box models via machine learning. Here, we show that the proposed interpretable workflow, which integrates physics-informed feature engineering consisting of Pearson-Spearman double correlation analysis and principal component analysis, Optuna-based hyperparameter optimization for the Gradient Boosting Decision Tree model, and SHapley Additive exPlanations mechanistic interpretation, achieves remarkable predictive performance with a training set coefficient of determination of 0.9731, a test set coefficient of determination of 0.9117, without overfitting, and exhibits satisfactory generalization on six refractory multi-principal element alloy compositions which does not appear in the dataset. This interpretable model provides actionable compositional design guidelines, bridging the gap between data-driven modeling and experimental development via its high precision, interpretability, and small-dataset adaptability. This work not only provides a reliable tool for accelerating the rational design of high performance RMPEAs, but also establishes a reliable generalizable paradigm. Its modular structure enables it to predict other properties of RMPEAs. This may further facilitate the development of next-generation refractory materials for aerospace and energy applications.