TL;DR
Princeton's PACMAN framework ran a real tokamak's plasma controls in real time at 20-millisecond loops, predicting and preventing instabilities 200 ms before they formed, a capability no human operator can match and a potential prerequisite for commercial fusion.
What happened
- Princeton University and PPPL unveiled PACMAN (Prediction And Control using MAchiNe learning), an AI framework for real-time fusion plasma control, with results published in the journal Nuclear Fusion.
- Five separate experiments on the DOE's DIII-D National Fusion Facility tokamak in San Diego validated the system across distinct control tasks.
- PACMAN's control loop runs every 20 milliseconds, versus the seconds-scale reaction time of a focused human operator, and repeats continuously throughout an experiment.
- The framework predicted a tearing-mode instability 200 milliseconds in advance, allowing corrective action before the instability formed rather than after, which conventional controllers cannot do.
- PACMAN simultaneously managed all six gyrotrons on DIII-D, dynamically adjusting microwave output power and mirror positions to hit researcher-defined plasma targets.
Why it matters
- Plasma instabilities grow in milliseconds: traditional simulation runs take days to months, making real-time human or simulation-based control physically impossible at reactor scale.
- Modular architecture lets researchers add, remove, or swap individual ML models without rebuilding the platform, compressing the iteration cycle for fusion experiments significantly.
- A system that prevents instabilities rather than suppressing them avoids the performance degradation that comes with reactive suppression, directly improving plasma quality and experiment yield.
- The demo establishes AI not as an assistant to fusion operators but potentially as load-bearing infrastructure: without millisecond-scale autonomous control, some future reactor operating regimes may simply be unreachable.
- Human oversight is preserved by design: operators define objectives and hard safety limits; PACMAN enforces those limits at the output stage before any command reaches the tokamak hardware.
What to watch next
- Whether PPPL and partners port PACMAN to other tokamak platforms beyond DIII-D, which would confirm it as a general fusion-control infrastructure layer rather than a single-machine demo.
- Commercial fusion developers (TAE, Commonwealth Fusion, Helion, and others) adopting or licensing modular ML control frameworks as a signal that the technology is crossing from research to engineering.
- Progress on reinforcement-learning models taking full autonomous control of heating systems, which PACMAN already demonstrated in one experiment, as a leading indicator of how much human oversight will shrink over time.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.