Interpretable Machine Learning Analysis of Reflection Loss in High-Entropy Electromagnetic Materials

A physics-informed machine learning framework was developed to predict reflection loss (RL) behavior in high-entropy electromagnetic absorbers using a literature-derived dataset. Frequency, thickness, density, attenuation constant, magnetic loss tangent, average atomic radius, average electronegativity, and valence electron concentration (VEC) were used as input descriptors. Among the investigated models, the neural network achieved the highest prediction accuracy (R² = 0.806), followed closely by gradient boosting (R² = 0.804). SHAP and feature-importance analyses revealed that average atomic radius and electronegativity were the dominant factors governing RL behavior. Larger atomic-radius values promoted lattice distortion and dielectric relaxation, whereas electronegativity mismatch enhanced charge redistribution and dipole polarization behavior. Thickness and frequency additionally influenced RL through quarter-wavelength resonance conditions.