Machine Learning Enabled Models to Predict Sulfur Solubility in Nuclear Waste Glasses

Abstract

The U.S. Department of Energy is considering implementing the direct feed approach for the vitrification of low-activity waste (LAW) and high-level waste (HLW) at the Hanford site in Washington state. If implemented, the nuclear waste with a higher concentration of alkali/alkaline-earth sulfates (than expected under the previously proposed vitrification scheme) will be sent to the vitrification facility. It will be difficult for the existing empirical models to predict sulfate solubility in these glasses or design glass formulations with enhanced sulfate loadings in such a scenario. Further, the existing models are unable to produce reliable predictions when applied to HLW glasses whose composition falls outside of the range encompassed by the database used to develop/calibrate the models. Accordingly, this study harnesses the power of artificial intelligence (machine learning, ML) with a goal to address the limitations of the existing models. Toward this, three ML models have been trained using a large database; comprising >1000 LAW and HLW glasses and encompassing a wide range of glass compositions and processing variables. Next, the ML model with the best prediction performance has been used to quantitatively assess and rank the influence (i.e., importance) of glasses’ compositional/processing variables on the SO3 solubility in the glasses. Finally, on the premise of such understanding of influential and inconsequential variables, two closed-form analytical modelswith disparate degrees of complexity (one highly parametrized and one with fewer input variables)have been developed. Results show that both analytical models produce predictions of SO3 solubility in LAW and HLW glasses with an accuracy analogous to ML models and substantially higher than the analytical models that represent the current state-of-the-art. Overall, this study’s outcomes present a roadmapinformed by data and channeled by artificial intelligencethat can be leveraged in the future to design nuclear waste glasses with unprecedented sulfur loadings.

A Deep Learning Approach to Design and Discover Sustainable Cementitious Binders: Strategies to Learn From Small Databases and Develop Closed-form Analytical Models

“…As can be seen, this model unites the fast Fourier transformation (FFT) algorithm with the deep learning (DL) model. Details of the DL model-which is premised on the random forests model that has been in our previous studies (Cook et al, 2021b;Lapeyre et al, 2021;Xu et al, 2021)-can be found in Supplementary Section S1 of Supplementary Information S1…”

“…The FT-DL model described in section 2.0 differs from the ML models used in our previous studies (Cook et al, 2021b;Lapeyre et al, 2021;Xu et al, 2021) (i.e., DL model based on random forests) in one key respect: In the FT-DL model, the database is FFT-transformed, prior to the model's training, so as to reduce the database's dimensionality; whereas in the DL model, the database is used in its pristine form. In section 2.0, it was argued that the FFT-transformation of the database ensures better training of FT-DL model; thereby, resulting in improvement of its prediction performance.…”

“…To develop a reliable analytical model, it is crucial to select input variables that cast significant influence on the output, while disregarding those which are largely inconsequential. The "DL" part of the FT-DL model is important in this context; because, it can statistically evaluate-in the form of Gini scores (Xu et al, 2021;Han et al, 2020a;Han et al, 2020b;Breiman, 2001;Menze et al, 2009)-the influence of each variable on the output. Results from this analysis are shown in Figure 8.…”

Impacts of temperature on sulfur solubility in low‐activity waste glasses

“…Instead, crucible-scale sulfur solubility tests have been used to augment melter test results to develop a compositional model. 14,[26][27][28][29][30][31] The model approach includes predicting sulfur solubility by the compositional model and then calculating melter tolerance values based on the correlation between the sulfur solubility and the melter tolerance. However, it is not clear how the two values are correlated.…”

“…Due to the cost and time required to perform melter tests, the directly measured melter tolerance data are limited. Instead, crucible‐scale sulfur solubility tests have been used to augment melter test results to develop a compositional model 14,26–31 . The model approach includes predicting sulfur solubility by the compositional model and then calculating melter tolerance values based on the correlation between the sulfur solubility and the melter tolerance.…”