August 7, 2026
By:
Lili Sarajian
Fusion has enormous potential to address the world’s energy crisis, but the high upfront capital investment required to design and build a fusion power plant is a barrier to commercializing this clean energy source.
One of the main cost drivers for fusion power plants is plasma confinement, or how well the device can contain the ultrahot, electrically charged gas within. Proper confinement maintains the conditions needed for fusion reactions to occur and produce energy.
In turn, the level of plasma confinement required is largely determined by plasma turbulence—small-scale density and temperature fluctuations in the plasma. Specialized diagnostic instruments on fusion devices capture the fundamental plasma dynamics behind these fluctuations.
Real-time analysis of that data could enable the system to automatically and instantaneously “course correct,” maintaining the optimal plasma state. However, analyzing that data in real time requires high-throughput computational capabilities at the limit of current technologies.
David Smith, a senior scientist in the Department of Nuclear Engineering and Engineering Physics at the University of Wisconsin–Madison, is leading a project that will use AI and machine learning to make this real-time plasma control possible.
With funding from the U.S. Department of Energy (DOE) Genesis Mission, Smith’s team will develop a first-of-its-kind, high-throughput edge AI to provide critical and previously unavailable information about the plasma state for autonomous control. This is one of five projects led by UW–Madison researchers that was awarded through the Genesis Mission.
Continuing UW–Madison’s longstanding collaboration with the DIII-D tokamak, a DOE user facility hosted by General Atomics, the project will use data from a fluctuation diagnostic developed and operated by the university. The diagnostic system produces a high-bandwidth data stream approaching a gigabyte of data per second.
“Analyzing that data in real time and then producing an inference calculation is technically challenging,” says Smith. “It cannot be done without AI.”
To overcome the challenges of processing large streams of data, this work will employ edge AI. In this method, AI calculations are performed near the diagnostic instrument. So, rather than moving the data from the sensor to a centralized computer, this edge AI technique deploys a small, specialized processor near the sensor.
As it is, the data can’t be processed by a standard CPU or GPU. It requires an integrated circuit known as a field programmable gate array, augmented with neural networks. Neural networks are a kind of machine learning model inspired by neurons in the human brain that process information in progressive layers, detecting patterns from the raw data.
This process will generate real-time inferences about fluctuations in the plasma that are sent to a plasma control system. The plasma control system interprets those signals, then commands actuators to adjust elements of the system—like heating, magnets and fueling—steering the plasma to the desired state.
Constrained by existing capabilities, most fusion devices operate in short pulses, and the diagnostic data is analyzed after each “shot.” Commercially viable fusion devices must shift to quasi-continuous and quasi-autonomous operation to maximize energy production and lower costs. High-throughput edge AI technology will make that possible.
Smith will collaborate with UW–Madison scientist Semin Joung and Ryan Coffee, a researcher at SLAC National Accelerator Facility who will lend his expertise in edge AI for high bandwidth data streams.
Their machine-agnostic approach has promising implications for plasma control in a wide range of fusion machines.
“Our vision is to bring together measurements of fundamental plasma dynamics and the real-time processing of sensor data to control fusion plasmas in a manner that is independent of machine-specific behavior,” says Smith. “This will only be possible by leveraging AI and machine learning models on specialized computer hardware.”
Featured image caption: The DIII-D National Fusion Facility where the University of Wisconsin–Madison developed and operates a diagnostic on the tokamak to capture fundamental plasma dynamics. Image courtesy of General Atomics.