Machine Learning Enabled Prediction and Interpretation of the Initial Dissolution Rates of Nuclear Waste Glasses

Abstract

Accurate prediction of chemical durability, such as dissolution rate, is crucial for glass materials used in diverse applications, including pharmaceutical packaging and nuclear waste disposal. In this work, machine learning‐based prediction models were developed using structural descriptors derived from molecular dynamics (MD) simulations and experimental conditions (such as pH and temperature). Structural features such as cation coordination, glass‐former cation polyhedral linkage density, and non‐bridging oxygen ratio were obtained from analyzing glass structure models generated from MD simulations using effective potentials. The dataset comprised 90 initial dissolution‐rate measurements for borosilicate and aluminosilicate glasses, covering static tests and dynamic stirred reactor coupon analysis tests (SRCA, ASTM C1926‐23) of International Simple Glass (ISG) and its direratives. The estimated coefficients suggest that the Ridge regression model exhibited better predictive performance, with the training dataset yielding a coefficient of determination ( R 2 ) of 0.8777 and a root mean square error (RMSE) of 0.7341. In addition, a non‐linear Random Forest model was employed, which showed comparable predictive capability, with R 2 = 0.8563 and RMSE = 0.7262. Strong generalization was achieved across multiple sources ( R 2 = 0.8210–0.9596) from leave‐one‐group‐out cross‐validation, which treated each source as an independent test set. It was found that dissolution rate was strongly associated with bond‐breaking propensity during initial surface hydrolysis, with former‐O‐Zr, Si‐O‐Si, Al‐O‐Si, Al‐O‐B, and Si‐O‐B linkages negatively correlating with dissolution rate hence improving chemical durability. These results demonstrate the potential of combining atomistic structural descriptors from MD simulations with machine learning methods to predict glass dissolution behavior.