Assessing static glass leaching predictions from large datasets using machine learning

Abstract Radioactive waste vitrified within glass is planned to be ultimately disposed of within a geological disposal facility. This study has applied machine learning to predict static glass leaching using an international experimental database of approximately 450 glasses to train/test various algorithms. Machine learning can accurately predict B, Li, Na, and Si releases for this complex database with Tree-based algorithms (notably ‘BaggingRegressor’ and ‘RandomForestRegressor’ in Python). This is provided that leaching experiment results, including elemental releases, are incorporated within the algorithm training variables, given that this study finds inaccurate prediction solely using initial test parameters as features. The trained algorithms underwent additional testing using an external database with prediction showing worse performance, likely due to substantial MgO and Na2O pristine glass oxide compositional variations across databases, with B releases generally being overestimated and Na underestimated. The use of molar oxide content performed significantly better than weight-fraction oxide for learning.