Physics-informed, uncertainty-aware surrogate screening of gamma-ray attenuation in lead-free multicomponent oxide...

Lead-free multicomponent oxide glasses are attractive candidates for gamma-ray shielding because their attenuation behavior can be tuned through composition while avoiding the toxicity of conventional lead-containing materials. In this work, we develop a physics-informed, uncertainty-aware surrogate framework for the pre-experimental screening of gamma-ray shielding performance in the B₂O₃-TeO₂-GeO₂-Bi₂O₃-BaO-ZnO glass system. A composition-energy dataset was constructed over 0.015-15 MeV using XCOM-based elemental mass attenuation coefficients, mixture-rule calculations, and physically motivated compositional descriptors. The primary learning target was the mass attenuation coefficient (μ/ρ), while density-dependent quantities were treated as derived screening metrics through a composition-based density estimate. Because (μ/ρ) is deterministically obtained from elemental XCOM data through the mixture rule, the surrogate is used here as a reproducible methodological framework for grouped validation, uncertainty calibration, interpretability, and screening-oriented ranking rather than as a replacement for the direct calculation. To avoid composition leakage, model development and evaluation were performed using grouped splits defined at the glass-composition level. Among the tested regressors, the random forest model provided the most accurate surrogate for (μ/ρ), achieving R²=0.9996 and a mean absolute percentage error of 0.742% on previously unseen compositions. An ensemble-based calibration strategy yielded 95% prediction intervals with empirical coverage of 0.950 on the unseen test set, enabling uncertainty-aware comparison of candidate glasses rather than point-estimate ranking alone. Interpretation analyses consistently identified photon energy as the dominant control variable and Bi₂O₃-related descriptors as the leading compositional contributors to enhanced attenuation, which provides a physics-consistency check confirming that the surrogate recovers the expected energy and high-Z dependence of photon attenuation rather than indicating a new attenuation mechanism. The framework was then used to screen 5000 synthetically generated candidate glasses at 0.1 MeV and to identify a small set of compositionally plausible, high-performing candidates for follow-up evaluation. Rather than replacing physics-based transport calculations or experiment, the proposed workflow provides a reproducible surrogate-screening strategy for composition-space exploration, interpretable trend analysis, and uncertainty-aware prioritization in the design of lead-free gamma-ray shielding glasses.