Abstract
Over 200 million liters of nuclear waste is stored in tank farms in Hanford, Washington, USA. An ongoing effort to immobilize non-solid nuclear waste is through waste vitrification, where waste is mixed and melted with glass-forming chemicals (GFCs) to form a waste glass. Algorithms to determine the ideal mixing recipes are formalized as constrained optimization problems to maximize waste loading in the glass product, with constraints on several properties of the glass melt and final product. These properties are on the glass melt, such as viscosity, electrical conductivity, and corrosion rates, as well as the final glass product, such as as chemical durability and retention of various leachants. The research presented in this dissertation uses contemporary data science methods to improve the formulation of nuclear waste glass, reducing the total mass of glass produced over the lifetime of the vitrification operation, saving time and resources. These methods are machine learning (ML), uncertainty quantification (UQ), and gradient-based optimization. ML models are trained on glass property data to replace statistical models as more accurate data-driven models and interfaced into an optimization platform to expand the feasible space and processing envelope of glass vitrification algorithms. Approximate UQ methods are developed and integrated in optimization alongside ML models for trustworthy solutions. Parametric ML methods like neural networks and approximate uncertainty propagation (UP) methods can be much faster and as accurate as alternative methods. The primary contributions of this effort are stated below. First, a survey of open source modeling platforms and packages, as well as future directions, is performed. The resulting contributions of this dissertation are built from open source tools developed by the scientific computing community. Second, ML models, such as Gaussian process regression, neural networks, support vector regression, and tree-based models are interfaced into a gradient-based optimization platform for surrogate model optimization. Interfaced models can be used as constraints or objective functions for optimization in a symbolic framework. By integrating ML models into optimization, vitrification algorithms are improved with access to more accurate ML models and direct optimization by gradient-based methods that rely on the Karush-Kuhn Tucker conditions. While offering more accurate and trustworthy property predictions, this integration improves an example low activity waste formulation solution by increasing the maximal waste loading from 34 wt % to 37.5 wt % over the baseline, and improves average waste loading for full optimization of 20 characteristic high level waste clusters from 37 wt % to 37.6 wt % over baseline. Third, methods of uncertainty quantification and propagation are integrated within vitrification algorithms alongside more accurate ML models. Assumed input distributions are typically propagated through a system with Monte Carlo sampling resulting in output distributions, which can be used for confidence intervals used as buffers from constraints. Analytical methods of input uncertainty propagation, such as first order error propagation and surrogate modeling with ML, can produce uncertainty intervals equivalent to standard sampling approaches in a fraction of the time by eliminating excess iterations, offering 10 times faster solutions. Uncertainty intervals for model predictions can be produced by model specific methods, ensemble methods, conformal prediction, or the Delta method. Pairing UQ alongside ML allows black-box models to be more trustworthy and reliable.
Degree
PhD
College and Department
Ira A. Fulton College of Engineering; Chemical Engineering
Rights
https://lib.byu.edu/about/copyright/
BYU ScholarsArchive Citation
Gunnell, LaGrande L., "Improvement of Hanford Waste Vitrification Algorithms With Machine Learning, Uncertainty Quantification, and Gradient-based Optimization" (2025). Theses and Dissertations. 10778.
https://scholarsarchive.byu.edu/etd/10778
Date Submitted
2025-05-01
Document Type
Dissertation
Handle
http://hdl.lib.byu.edu/1877/etd13582
Keywords
Nuclear Waste, Vitrification, Machine Learning, Uncertainty Quantification, Constrained Optimization
Language
english