Machine learning to predict refractory corrosion during nuclear waste vitrification

Abstract

The goal of this study was to determine the effects of model nuclear waste glass composition on the corrosion of Monofrax® K-3 refractory, using machine learning (ML) methods for data investigation and modeling of published borosilicate glass composition data and refractory corrosion performance. First, statistical methods were used for exploration of the data, and the list of features (model terms) was determined. Several model types were explored, and the Bayesian Ridge type was the most promising due to low mean average error and mean standard error as well as high R2 value. Parameters and model results using previously identified model features and those from this study are compared. ML methods appear to give results at least as good as previously available models for describing the effects of glass composition on refractory corrosion.