Predicting Nepheline Formation in Nuclear Waste Glasses Using Machine Learning With Uncertainty Quantification

ABSTRACT The vitrification of high‐level waste (HLW) into borosilicate glass is a cornerstone of nuclear waste management, yet one of the key challenges in glass formulation for high‐aluminum‐containing waste streams is the precipitation of nepheline during the canister cooling stage. Excessive nepheline formation compromises glass durability, threatening the long‐term stability of vitrified waste. Thus, predictive models capable of accurately anticipating nepheline formation from glass composition are critical for waste glass formulation and optimizing stability and durability. Various methods such as discriminators and sub‐mixture models have been developed to determine nepheline precipitation; however, they tend to be overly conservative and unnecessarily limit achievable glass compositions. Building on previous works, we explore the application of several classical machine learning models—Gaussian Processes, Gradient Boosting, and Random Forest—alongside Bayesian methods, including fully Bayesian Gaussian Processes and Partial Bayesian Neural Networks, to address the challenges posed by limited datasets and prediction uncertainties intrinsic to this problem. Furthermore, we introduce a dual‐head partially Bayesian neural network with scaled regression loss, validated through ablation, to directly predict nepheline precipitation. Notably, this regression‐based strategy enables the use of practical nepheline tolerance thresholds, substantially reducing false positives compared with existing binary nepheline discriminator approaches that cannot distinguish between trace and significant nepheline formation—expanding access to novel compositional regions for waste glass design. By balancing explainability, uncertainty quantification, and accuracy, our results lay the foundation for more efficient optimization of HLW vitrification processes and contribute to the broader goal of improving nuclear waste management strategies.