Predicting the Long-Term Durability of Nuclear Waste Immobilization Glasses Using Machine Learning

The long-term durability of a glass is a key performance metric for nuclear waste immobilization application. Although some models have been proposed to predict the short-term forward dissolution kinetics of glasses, long-term dissolution is a more complex behavior that is influenced by the glass structure, the feedback from the solution, and the precipitation of secondary phases. This complexity has limited our ability to robustly predict the long-term dissolution rate of nuclear waste immobilization glasses. Here, based on the analysis of a large dataset of vapor hydration tests (VHT), we develop using machine learning a Gaussian Process Regression (GPR) model to predict the long-term durability of glasses. Importantly, GPR models are non-parametric and intrinsically capture the uncertainty of the prediction. We demonstrate that our GPR model features an excellent accuracy. This model allows us to decipher the propensity for each oxide to accelerate or slow down the dissolution kinetics of glasses.