Uncertainty-aware machine learning with conformal prediction for gamma-ray attenuation performance in...

In this study, we developed and evaluated an uncertainty-aware machine-learning model for the gamma-ray attenuation of 75TeO₂-5Na₂O-(20-x)BaO-xTiO₂ glasses (x = 0-20 mol%) and used split conformal prediction to quantify prediction uncertainty. Mass attenuation coefficient (MAC) values were calculated with NIST XCOM for 21 compositions and 100 photon energies from 0.001 to 100 MeV, giving 2100 composition-energy records. The data were divided into development (70%), calibration (15%), and independent test (15%) subsets. An initial screen of more than 50 regression configurations was followed by a final comparison of eight representative model families using random-row, held-out-composition, and held-out-energy cross-validation. Extra Trees gave the lowest held-out-energy error and was selected. On the 315-row test subset, the model reached a coefficient of determination (R²) of 0.999999, a root mean square error (RMSE) of 0.001869 in log₁₀(MAC), a mean absolute percentage error (APE) of 0.165%, and a maximum APE of 3.446%. The two interpolation baselines gave mean APEs of 3.923% and 3.988%, while direct mixture calculation from cached XCOM elemental tables reproduced the reference values to numerical precision and was faster than machine-learning inference. Energy-stratified split conformal prediction achieved 91.75% empirical coverage for a 92% nominal target. The largest test errors occurred near low-energy absorption-edge regions. The prediction intervals quantify surrogate error relative to XCOM and should not be interpreted as experimental uncertainty.