Machine learning screening of magnetocaloric high-entropy alloys using a field-normalized ΔSref target

High-entropy alloys (HEAs) have emerged as promising candidates for magnetocaloric applications due to their unique structural and magnetic properties. Finding the right HEA for next-generation magnetic cooling technologies is a complex puzzle. Our work introduces a machine-learning framework to address this by predicting a field-normalized magnetic entropy change (?Sref) from an alloy's chemical composition, with screening performance benchmarked against measured ? S M at the reported field for practical comparability. To ensure robustness, we rigorously evaluated five regression algorithms using a nested cross-validation strategy, and tree-based models proved exceptionally effective. The LightGBM model emerged as the top performer for ?Sref prediction, achieving a mean R2 = 0.801 ± 0.212. We used SHAP analysis to understand which features drive these predictions, revealing that ?Sref is primarily informed by mixing enthalpy (?Hmix), mixing entropy (?Smix), and valence electron concentration (VEC). These insights allowed us to establish a clear set of composition-based design rules for identifying promising alloys with high ? S M (? 9 J kg?1 K?1). From pooled log–log scaling (? S M ? H?), we obtain n ? 0.91; at a conservative operating point the composition-only rule attains coverage = 0.347, recall = 0.720, and precision = 0.261. We further confirm generalization via external (literature-based) validation on 15 unseen compositions (11/15 correct; ?0.73, 95% Clopper–Pearson CI 0.45–0.92), supporting the utility of the proposed framework for pre-experimental compositional screening of magnetocaloric HEAs. © 2026 Elsevier B.V.