PNNL Uses Machine Learning to Reduce Nuclear Waste Volume in Glass Storage
Scientists at Pacific Northwest National Laboratory have used machine learning to optimize glass formulas for storing radioactive waste, potentially reducing the number of containers needed by 5% over the life of the Hanford cleanup mission.
The research, published in the Journal of Non-Crystalline Solids in April, addresses a longstanding problem: Hanford Site tank waste contains nearly every element on the periodic table in varying concentrations, making it chemically the most complex radioactive waste in the world. Traditional methods struggle to find glass formulations that safely incorporate the maximum amount of waste while meeting durability and processing requirements.
PNNL researchers trained machine learning models on decades of data from Hanford waste samples to predict which combinations of waste and additives would allow higher waste loading - the percentage of radioactive material incorporated into each glass container. The models tested thousands of combinations that human scientists hadn't previously considered.
The Waste Problem
More than 50 million gallons of radioactive waste sits in underground tanks at Hanford, much of it generated during plutonium production in the Manhattan Project and Cold War eras. The waste composition changes over time and varies tank to tank, even within individual tanks, as chemical reactions occur and material is transferred during treatment.
Current operations use a 2012 algorithm that intentionally accepts lower waste percentages to reduce variables during processing. Xiaonan Lu, a materials scientist who led the new study, said the machine learning approach works differently: "Models can learn from their own mistakes. We replaced a traditional math equation with a machine learning model that tried every combination of elements that have been measured in the Hanford tank waste samples."
Glass as a Disposal Method
The Hanford Waste Treatment and Immobilization Plant converts liquid waste into solid glass by mixing waste with additives, heating the mixture to 2,100°F, and pouring it into 7-foot-tall steel containers. The process requires precise control - glass that's too thin risks corroding the melter; glass that's too thick won't pour properly.
Current low-activity waste glass typically holds 20%-30% radioactive material by weight. The new models show waste loading can increase by roughly 1% for every 20% already in the formula, reducing the total volume of glass needed for disposal.
José Marcial, a materials scientist at PNNL, explained the operational impact: "A 5% reduction in glass logs would have ripple effects for the mission's operational timeline, number of containers needed, and storage space in the Hanford Site's Integrated Disposal Facility."
Broader AI Applications in Cleanup
The work aligns with the DOE's Genesis Mission, launched in November 2025, which identified 26 national science challenges where AI could accelerate research. "Transforming nuclear restoration and cleanup" was one challenge.
Four PNNL researchers are part of the Genesis Mission's Nuclear Restoration and Revitalization AI-Roadmap team, working with other national labs to identify opportunities to use AI for environmental cleanup across complex sites.
Matt Asmussen, senior materials scientist at PNNL, said the glass research demonstrates the broader potential: "We're combining PNNL's decades of expertise in glass science and vitrification with advanced AI tools to compress timelines and give a compelling preview of what mission acceleration could look like through the use of AI/ML models."
The research represents the first experimental validation of active learning in waste glass design, according to John Vienna, a Lab Fellow who was part of the original 2012 algorithm team. "It's the last 14-year journey from the original glass algorithm application to now that has been the most exciting," Vienna said.
Learn more about AI for Science & Research and how organizations apply machine learning to complex scientific challenges.
Your membership also unlocks:
