For more than seven decades, the Hanford Site in Washington State has stored some of the most chemically complex radioactive waste in the world. During the Manhattan Project and the Cold War, operators at Hanford generated massive quantities of plutonium production waste, which was then stored in massive underground tanks. Over 50 million gallons of this waste remains on site, a legacy of America's nuclear weapons program that has become one of the country's most challenging environmental remediation projects.

The waste is extraordinarily complex. Nearly all elements on the periodic table exist in Hanford tank waste, creating a chemical mixture of staggering complexity. The composition varies not just from tank to tank, but within each tank and as the waste is transferred during treatment. For decades, scientists and engineers have struggled to find an efficient way to immobilize this waste—to convert it into a stable, long-term storage form.

The traditional approach has been vitrification: heating the waste with glass-forming additives to 2,100 degrees Fahrenheit and pouring the molten mixture into 7-foot-tall steel containers. The resulting glass is stable and can be stored safely for thousands of years. But the challenge lies in optimization. The waste must be mixed with precisely the right combination of additives to ensure that the resulting glass meets durability requirements, flows properly through the treatment plant, and maximizes the amount of waste incorporated into each container.

Machine Learning Breakthrough

This is where artificial intelligence enters the story. Scientists at the Department of Energy's Pacific Northwest National Laboratory (PNNL) have developed machine learning models that can predict optimal glass formulations for the complex Hanford waste. The models were trained on decades of experimental data and can rapidly evaluate thousands of potential combinations of waste properties and additives to identify the formulations that maximize waste loading—the percentage of waste incorporated into the final glass—while ensuring that the glass meets all technical requirements.

The results have been striking. Using traditional approaches, the PNNL team's original algorithm, developed in 2012, was intentionally conservative, accepting lower quantities of waste mixed with additives to reduce variability and simplify the treatment process. The new AI-driven models can increase waste loading by approximately 1 percent for every 20 percent of waste already going into the recipe. Over the life of the vitrification project, this improvement could reduce the number of glass logs produced by 5 percent—a seemingly modest improvement that translates into enormous cost savings and reduced environmental impact.

The Hanford Site's vitrification project is expected to run for decades and cost tens of billions of dollars. Every reduction in the number of glass containers produced translates directly into reduced operational costs, reduced disposal facility footprint, and accelerated project completion. A 5 percent reduction in glass logs produced could save hundreds of millions of dollars and decades of operational time.

"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. That's decades of data the model used to learn and then predict which recipes would work."

— Xiaonan Lu, Materials Scientist, PNNL