Executive Overview
For decades, the pursuit of commercial nuclear fusion—the very process that powers the sun and stars—has been heralded as the holy grail of clean energy. If successfully harnessed, fusion promises a virtually limitless, safe, and carbon-free supply of electricity, capable of satisfying global energy demands for millennia without the long-lived radioactive waste or catastrophic meltdown risks associated with traditional fission reactors. Yet, despite monumental scientific advances, the realization of practical fusion has remained perpetually out of reach, frustrated by a single, unforgiving constraint: physics operates on a timescale too fast for human intervention.
Inside experimental magnetic confinement reactors known as tokamaks, hydrogen isotopes are heated to extreme temperatures exceeding 100 million degrees Celsius—hotter than the solar core—transforming them into an electrically charged, swirling soup of plasma, often referred to as the fourth state of matter. To prevent this ultra-hot plasma from melting the reactor walls, scientists confine it using immensely powerful magnetic fields. However, this plasma is inherently unruly. Infinitesimal disturbances, thermal fluctuations, or minor magnetic imbalances can cascade into catastrophic instabilities within mere milliseconds. These rapid deviations can quench the fusion reaction or damage the physical infrastructure of the machine.
Human operators, no matter how skilled or attentive, possess reaction times measured in seconds. They are biologically and technologically blind to crises that unfold in fractions of a blink of an eye.
To bridge this fatal temporal gap, a collaborative team of researchers at the U.S. Department of Energy’s (DOE) Princeton Plasma Physics Laboratory (PPPL) and Princeton University has engineered a groundbreaking software framework called PACMAN (Prediction And Control using MAchiNe learning). Published recently in the prestigious journal Nuclear Fusion, PACMAN represents a paradigm shift in how artificial intelligence interfaces with heavy physics hardware. By deploying a modular, high-speed machine learning ecosystem that operates on a continuous 20-millisecond loop, PACMAN successfully predicts, intercepts, and corrects plasma instabilities before they manifest, while strictly adhering to hardware safety protocols and keeping human scientists firmly at the helm of strategic objectives.
Following a successful battery of five validation experiments at the DOE’s DIII-D National Fusion Facility in San Diego, PACMAN has proven that AI can transform fusion control from disjointed, one-off maneuvers into cohesive, real-time infrastructure.
Detailed Chronology: From Concept to DIII-D Validation
The genesis of PACMAN was born out of a stark realization within the plasma physics community: while advanced computational physics simulations are exceptionally detailed, they are fundamentally useless for real-time control. Traditional first-principles physics models can take hours, days, or even months of supercomputer time to compute the trajectory of a plasma discharge. While invaluable for post-experiment analysis and long-term reactor design, these legacy models cannot assist an operator running a multi-million-dollar experiment that lasts only a few minutes.
Recognizing this bottleneck, graduate students and lead authors Hiro Farre Kaga (Princeton Program in Plasma Physics) and Andy Rothstein (Department of Mechanical and Aerospace Engineering) set out to build an architecture capable of processing live telemetry and executing corrective actions at machine speeds.
Phase 1: Conceptualization and Modular Architecture
Historically, machine learning in fusion energy suffered from a lack of integration. Individual research teams would train isolated neural networks to perform narrow tasks—such as predicting a specific magnetic perturbation or adjusting a single heating beam—without a standardized communication protocol. When deployed on a physical tokamak, these disparate models often clashed, lacked synchronization, or introduced critical software latencies.
The Princeton and PPPL team approached the problem by conceptualizing PACMAN not as a monolithic algorithm, but as a modular assembly line divided into four distinct, highly optimized stations:
- Data Ingestion and Sanitization: The framework instantly captures live sensor measurements from the tokamak—including high-frequency magnetic signals, core temperatures, and plasma density profiles—filters out noise, and packages the data.
- Predictive Estimation: Dedicated machine learning models ingest the sanitized data packages to evaluate current plasma states and forecast future behavior over millisecond horizons.
- Controller Execution: Mathematical controllers evaluate the predictions and determine the requisite physical adjustments, such as modulating auxiliary heating systems or shifting magnetic coil currents.
- Safety Arbitration and Dispatch: The final stage acts as an unyielding gatekeeper. It resolves conflicting commands from multiple controllers, applies hardcoded hardware safety limits, and transmits the finalized, verified instructions to the tokamak actuators.
Phase 2: Experimental Deployment at DIII-D
With the framework architected, the team traveled to the DIII-D National Fusion Facility in San Diego, California, operated by General Atomics for the DOE Office of Science. DIII-D is one of the world’s premier magnetic fusion research devices, providing an ideal, highly instrumented testbed for advanced control algorithms.
Over a rigorous campaign comprising five distinct experimental runs, the researchers put PACMAN through its paces. The trials were designed to test the framework’s ability to handle multi-variable optimization, preempt rapid instabilities, and integrate multiple concurrent neural networks.
The results exceeded expectations. Most notably, in experiments targeting tearing mode instabilities—localized disruptions of the magnetic field lines that degrade confinement and typically force operators into reactive damage control—PACMAN’s predictive models flagged the impending anomaly up to 200 milliseconds before it occurred. Rather than reacting to a degraded plasma state, the framework dynamically altered the plasma parameters to avert the tearing mode entirely, preserving optimal thermal confinement and experimental performance.
Furthermore, PACMAN successfully orchestrated all six of DIII-D’s gyrotrons—powerful microwave beam injection systems used to heat the plasma. Operating in unison, the AI framework dynamically adjusted the power output and continuously repositioned the physical mirrors of all six gyrotrons in real time to meet complex, pre-programmed experimental trajectories. Post-experiment data analysis confirmed that the system achieved an optimal multi-variable solution that no human operator or legacy algorithm had previously been able to calculate on the fly.
Supporting Context & Metrics
To fully appreciate the technological leap represented by PACMAN, one must examine the operational parameters, computational speeds, and structural efficiencies that define modern magnetic confinement fusion.
The Millisecond Dilemma: A Quantitative Comparison
| Metric / Parameter | Human Operator Baseline | Legacy Control Systems | PACMAN AI Framework |
|---|---|---|---|
| Response Time | 200 milliseconds to seconds | 50 to 100 milliseconds | ~20 milliseconds (continuous loop) |
| Predictive Horizon | Reactive (post-instability) | Immediate reactionary bounds | Up to 200ms advance warning (e.g., tearing modes) |
| Actuator Coordination | Serial / Single-system focus | Fragmented, single-purpose algorithms | Simultaneous orchestration (e.g., 6 gyrotrons) |
| Iteration / Integration Time | Months of custom recoding | Weeks of debugging per model | Days (second model integrated in ~48 hours) |
The Physics of Tokamak Confinement
Tokamaks utilize toroidal (doughnut-shaped) chambers surrounded by massive electromagnets. The primary toroidal magnetic field, combined with a poloidal magnetic field generated by an internal electrical current flowing through the plasma, creates a twisted magnetic cage. This cage is designed to keep charged particles from touching the interior walls, which would instantly cool the plasma and damage the vessel.
However, plasma is notoriously dynamic. As pressure and temperature rise to levels required for net energy gain ($Q > 1$), internal pressure gradients trigger magnetohydrodynamic (MHD) instabilities. These include edge-localized modes (ELMs), resistive wall modes, and neoclassical tearing modes. If left unchecked for even 10 to 50 milliseconds, these instabilities lead to "disruptions," where the plasma loses its thermal energy in a flash, dumping millions of joules onto the reactor walls and ending the experiment prematurely.
Scaling Software Agility
Beyond raw speed, PACMAN’s most profound metric of success lies in its software development lifecycle. Developing the framework and integrating the initial machine learning model required months of painstaking collaborative engineering. However, when the research team sought to integrate a second, distinct AI model into the operating loop, the modular architecture proved its worth: the integration process took a mere two days, exhibiting significantly fewer bugs and vastly simplified testing procedures. This dramatic reduction in friction enables rapid iteration cycles, allowing physicists to retrain, swap, and deploy novel algorithms weekly rather than annually.
Official Statements and Expert Insights
The implications of the PACMAN framework extend far beyond the immediate success of the DIII-D experiments, signaling a philosophical and technical maturation in how artificial intelligence is deployed within high-stakes scientific infrastructure.
Reflecting on the temporal gulf between human biology and plasma dynamics, co-lead author Hiro Farre Kaga emphasized the indispensability of machine learning models:
"That’s great for preparing for the next experiment in a year, but for control we need models that make a decision in the moment. Machine learning models can describe the plasma behavior very well, and importantly, they are the only way we have to model the plasma in millisecond times. The speed of these models is what’s key for control."
Detailing the collaborative mechanics of the software framework, co-lead author Andy Rothstein highlighted the qualitative leap from human reflexes to automated loops:
"We developed this framework so that models could communicate, outputs from those models could be shared and we could do exciting physics in one integrated system. A really focused human operator can respond on the order of seconds. The whole PACMAN framework typically runs in about 20 milliseconds, and it’s not running once. It’s running again and again and again. It can see small things happening in the plasma and adjust in a way that a human would never be able to do."
Addressing the surprise velocity of model iteration, Rothstein noted:
"DIII-D is first and foremost a research machine, and sometimes things don’t work out the way you expected. If you can put a model on in a week, you can retrain it and put a new one on the week after. It allows for iteration that wasn’t possible previously."
Crucially, the research team is careful to contextualize the role of AI, pushing back against dystopian narratives of autonomous machines stripping human scientists of control. Farre Kaga underscored that human stewardship remains absolute:
"No matter how sophisticated your controllers, in the end it’s a human operator that sets the parameters for that control."
This philosophical boundary is hardcoded into PACMAN: regardless of what an AI model recommends, the framework’s arbitration layer enforces strict, immutable hardware safety limits, while human physicists review comprehensive telemetry post-shot to refine operational strategies.
Looking toward the broader fusion community, Egemen Kolemen, associate professor of mechanical and aerospace engineering at Princeton University—jointly appointed with the Andlinger Center for Energy and the Environment and PPPL—summarized the overarching vision for PACMAN’s modular infrastructure:
"PACMAN uses a flexible setup where building-block AI algorithms can be put together. You can add a new one, swap one out or run several at once without touching the rest of the system. That modularity is what turns AI plasma control from a series of one-off demonstrations into infrastructure the whole fusion community can build on."
Future Outlook: Paving the Way for Commercial Fusion Reactors
The successful deployment of PACMAN at the DIII-D facility marks a crucial milestone, but it is merely the opening salvo in a much larger technological campaign. As the global fusion energy sector transitions from physics-exploratory science toward commercial engineering—exemplified by massive international undertakings like ITER and private ventures designing compact, high-field tokamaks—the demand for fault-tolerant, ultra-fast control systems will only intensify.
Adaptability Across Reactor Geometries
One of the most promising attributes of PACMAN’s design is its hardware-agnostic modularity. Because individual AI models, estimators, and controllers operate as isolated building blocks communicating through standardized data pipelines, the framework is not permanently tethered to the specific physical dimensions or magnetic configurations of DIII-D.
Princeton researchers and their collaborators—including contributors from Japan’s National Institutes for Quantum Science and Technology (QST)—are already eyeing the adaptation of PACMAN for other experimental devices. In principle, the framework can be scaled and tailored to tokamaks of varying shapes, sizes, and auxiliary heating suites, including next-generation spherical tokamaks and commercial pilot plant designs currently on the drawing boards.
Mitigating Divertor Heat Loads and Disruptions
Future iterations of PACMAN will target even more complex control challenges, such as managing extreme heat fluxes on divertor tiles (the components designed to exhaust waste heat from the reactor) and preempting major plasma disruptions altogether. By stitching together predictive models capable of forecasting thermal transport, impurity accumulation, and magnetohydrodynamic stability simultaneously, PACMAN could eliminate the catastrophic disruptions that have plagued magnetic confinement devices for half a century.
The Human-AI Symbiosis in Clean Energy
Ultimately, PACMAN redefines the relationship between human intellect and artificial intelligence in heavy scientific research. By offloading the terrifyingly fast, micro-second reflexes required to police a star-hot plasma onto a modular, unyielding software framework, human physicists are liberated to focus on what they do best: setting ambitious experimental objectives, interpreting complex physical phenomena, and designing the future of clean energy.
As the world stands on the precipice of a sustainable energy transition, innovations like PACMAN ensure that when commercial fusion reactors finally ignite, humanity will possess both the wisdom to guide them and the technological reflexes required to keep them burning.
