Scientists at the Department of Energy's Pacific Northwest National Laboratory (PNNL) have successfully utilized artificial intelligence to optimize glass formulas for treating liquid radioactive waste. Published in the April 15 edition of the Journal of Non-Crystalline Solids, the research demonstrates how machine learning can significantly increase the volume of waste incorporated into solid glass forms while reducing operational risks, costs, and mission duration. This advancement addresses the critical cleanup of legacy waste stored in massive underground tanks at the Hanford Site, which originated from plutonium production during the Manhattan Project and the Cold War. The waste is considered the most complex radioactive mixture in the world, containing nearly every element on the periodic table, including high concentrations of phosphorus and bismuth. Historically, vitrification—the process of immobilizing waste by converting it into glass—has faced challenges due to waste variability, complex chemistry, and strict safety requirements. The traditional mathematical algorithms developed in 2012 allowed for safe processing but intentionally accepted lower waste quantities to minimize variables. This conservative approach resulted in more glass containers being required, increasing the footprint of disposal facilities. To overcome these limitations, PNNL researchers replaced traditional equations with a custom machine learning model capable of analyzing decades of data from Hanford tank samples. This model learned from its own predictions and errors to identify optimal combinations of waste and chemical additives that would have been previously undiscoverable. According to Xiaonan Lu, a PNNL materials scientist and lead author, the AI model tested every measured combination of elements to predict which recipes would work most effectively. The research confirms that the new approach can increase waste loading, the percentage of waste within the final glass form. While traditional low-activity waste glass typically holds about 20% to 30% waste by weight, the new model suggests an increase of roughly 1% for every 20% of waste already in the recipe. Even a small percentage improvement has substantial impacts over the life of the project. John Vienna, a Lab Fellow and leading expert in the field, noted that a 5% reduction in the number of glass logs would significantly decrease the timeline, the number of containers needed, and storage space at the Integrated Disposal Facility. The optimization process ensures that the glass meets all durability and safety standards, much like the porridge in the Goldilocks story that must be neither too thin nor too thick. If the mixture is too thin, it risks corrosion within the melter; if too thick, it may not fill containers properly. The AI system simultaneously optimizes the formulation and physical properties to maintain efficient plant operations and ensure long-term stability for decades of storage. This project is part of the DOE's Genesis Mission, a broader initiative launched to use AI for transforming nuclear restoration and cleanup. Four PNNL researchers are currently leading the Nuclear Restoration and Revitalization AI-Roadmap team to operationalize these tools further. Matt Asmussen, a senior materials scientist, stated that this work highlights the strong potential of AI to compress timelines and accelerate complex cleanup missions by combining advanced machine learning with decades of glass science expertise. The study represents the first experimental validation of an active learning approach in waste glass design, offering a compelling preview of how artificial intelligence can transform nuclear remediation.