Public Abstract

DE-SC0026825: Weak-Form Agentic Digital Twins for Real-Time Control of Magnetically Confined Fusion Systems Under Sparse Observation

Award Status: Active

Institution: The Regents of the University of Colorado d/b/a University of Colorado, Boulder, CO
UEI: SPVKK1RC2MZ3
PM: Halfmoon, Michael

Most Recent Award Date: 09/22/2026
Number of Support Periods: 1
PI: Bortz, David

Current Budget Period: 09/01/2026 - 05/31/2027
Current Project Period: 09/01/2026 - 05/31/2027

Public Abstract

Weak-Form Agentic Digital Twins for Real-Time Control of Magnetically Confined Fusion Systems Under Sparse Observation University of Colorado Boulder: David M. Bortz (PI), Stephen Becker, Cristian López Los Alamos National Laboratory: Vitaliy Gyrya, Daniel A. Messenger Michigan State University: Andrew J. Christlieb Oak Ridge National Laboratory: Sebastian De Pascuale, Jeremy D. Lore Public Abstract Fusion energy, the process that powers the sun, could give the United States and the world a source of clean, abundant electricity. But making fusion practical on Earth requires exquisite real-time control of the ultra-hot ionized gas (plasma) inside a reactor. In particular, the plasma near a reactor’s inner walls can shift between stable and damaging states within milliseconds: too fast for human operators and too complex for existing controllers, especially since these devices carry only a handful of sensors, and those sensors are noisy. This project, led by the University of Colorado Boulder in partnership with Los Alamos National Laboratory, Oak Ridge National Laboratory, and Michigan State University, is part of the U.S. Department of Energy’s Genesis Mission to accelerate scientific discovery through AI. The team is developing a physics-constrained digital twin for the plasma boundary: a live AI model that runs alongside a fusion reactor, interprets sparse and noisy sensor data as it arrives, and issues control decisions fast enough to keep the plasma safe. The AI learns directly from live observations, discovering the equations that govern the plasma’s behavior and identifying the parameters that describe its current state; it remains explainable so operators can verify what the AI is doing, and honestly reports how confident it is in every prediction. The twin is built as an agentic system: a coordinating AI layer continuously monitors its own performance, detects when the plasma has entered a new regime, and autonomously decides whether to re-identify its parameters, re-discover its underlying equations, or tighten safety margins: all within the milliseconds available between control cycles. The team will demonstrate the twin on simulations of the DIII-D tokamak, a Department of Energy fusion facility in San Diego, showing that it outperforms conventional control approaches through the boundary transitions that future fusion pilot plants must handle safely. Beyond fusion, the same trustworthy agentic AI framework will extend to other high-stakes decision systems, from spacecraft to power grids, where reliable AI must operate on limited, noisy data.

Weak-Form Agentic Digital Twins for Real-Time Control of Magnetically Confined Fusion Systems Under Sparse Observation

University of Colorado Boulder: David M. Bortz (PI), Stephen Becker, Cristian López
Los Alamos National Laboratory: Vitaliy Gyrya, Daniel A. Messenger
Michigan State University: Andrew J. Christlieb
Oak Ridge National Laboratory: Sebastian De Pascuale, Jeremy D. Lore

Public Abstract

Fusion energy, the process that powers the sun, could give the United States and the world a source of clean, abundant electricity. But making fusion practical on Earth requires exquisite real-time control of the ultra-hot ionized gas (plasma) inside a reactor. In particular, the plasma near a reactor’s inner walls can shift between stable and damaging states within milliseconds: too fast for human operators and too complex for existing controllers, especially since these devices carry only a handful of sensors, and those sensors are noisy.

This project, led by the University of Colorado Boulder in partnership with Los Alamos National Laboratory, Oak Ridge National Laboratory, and Michigan State University, is part of the U.S. Department of Energy’s Genesis Mission to accelerate scientific discovery through AI. The team is developing a physics-constrained digital twin for the plasma boundary: a live AI model that runs alongside a fusion reactor,
interprets sparse and noisy sensor data as it arrives, and issues control decisions fast enough to keep the plasma safe. The AI learns directly from live observations, discovering the equations that govern the plasma’s behavior and identifying the parameters that describe its current state; it remains explainable so operators can verify what the AI is doing, and honestly reports how confident it is in every prediction.

The twin is built as an agentic system: a coordinating AI layer continuously monitors its own performance, detects when the plasma has entered a new regime, and autonomously decides whether to re-identify its parameters, re-discover its underlying equations, or tighten safety margins: all within the milliseconds available between control cycles.

The team will demonstrate the twin on simulations of the DIII-D tokamak, a Department of Energy fusion facility in San Diego, showing that it outperforms conventional control approaches through the boundary transitions that future fusion pilot plants must handle safely. Beyond fusion, the same trustworthy agentic AI framework will extend to other high-stakes decision systems, from spacecraft to power grids, where reliable AI must operate on limited, noisy data.