Nuclear Fusion

From PRIMO.ai
Jump to navigation Jump to search

Youtube search... ...Google search


Nuclear fusion is the reaction that powers the Sun and stars. Light atomic nuclei collide hard enough to overcome their electrical repulsion, fuse into a heavier nucleus, and release a large amount of energy. Most near-term power plant designs burn deuterium and tritium (D-T), two isotopes of hydrogen, because that reaction is the easiest to ignite. The catch is that the fuel has to become a plasma (a gas of free electrons and bare nuclei) hotter than 100 million °C, and it has to stay hot, dense, and confined long enough for more energy to come out than goes in.

Two broad strategies dominate the field:

  • Magnetic confinement fusion traps the plasma in strong magnetic fields. Tokamaks (doughnut-shaped machines such as DIII-D, JET, KSTAR, EAST, WEST, TCV, ITER, and Commonwealth Fusion Systems' SPARC) are the most mature design. Stellarators such as Wendelstein 7-X twist the magnetic field with complex external coils instead of relying on a large current inside the plasma. Companies are also pursuing alternatives, such as TAE Technologies' field-reversed configuration and Helion's pulsed magneto-inertial approach.
  • Inertial confinement fusion crushes a tiny fuel capsule with lasers or pulsed power so quickly that the fuel fuses before it can fly apart. Lawrence Livermore National Laboratory's National Ignition Facility (NIF) used this method to achieve the first laboratory ignition in December 2022.

This page covers where artificial intelligence (AI) and machine learning (ML) fit into that effort. The core contributors include Google DeepMind and the Swiss Plasma Center at EPFL (reinforcement learning control on the TCV tokamak); Princeton University and the U.S. Department of Energy's Princeton Plasma Physics Laboratory (PPPL), especially Egemen Kolemen's plasma control group and William Tang's deep learning disruption research; General Atomics, which operates the DIII-D National Fusion Facility; LLNL's cognitive simulation team at NIF; MIT's Plasma Science and Fusion Center working with Commonwealth Fusion Systems (CFS); and TAE Technologies, which has collaborated with Google Research since 2014. NVIDIA, Microsoft, and the DOE national laboratories supply much of the computing and data infrastructure.

The problem AI addresses. A fusion plasma is a nonlinear, turbulent, multi-scale system that physicists can only partly observe. The best first-principles simulations take days to months on supercomputers, yet the instabilities that end a plasma pulse can grow in a few thousandths of a second. Experiments are expensive and scarce: a large tokamak runs a few thousand pulses a year, and NIF fires only a small number of ignition-class shots annually. Future power plants will make this harder, because they will carry fewer diagnostics than today's research machines (sensors have to survive intense neutron bombardment).

The central mechanism. Researchers train fast models on simulation and experimental data, then use them in three ways: to predict (Will this plasma disrupt? Will this capsule ignite?), to estimate what sensors cannot directly see, and to decide (which coil voltages, heating powers, or operating scenarios to use). Physics codes, conventional controllers, and hard safety limits stay wrapped around the learned components.

The paradigm shift. Conventional plasma control is engineered loop by loop: separate controllers for plasma current, position, and each shape parameter, all designed around linearized models of one scenario at a time. Experiment planning traditionally follows a slow "run a shot, analyze for weeks, try again" cycle. AI changes both. A single learned policy can map raw sensor readings directly to every actuator, so physicists specify what they want rather than how to get it. Probabilistic models and digital twins let teams run predict-first experiments, where the model forecasts the result, with error bars, before the machine fires.

The idea is not new. In 2019, PPPL researchers reported that machine learning can help bring to Earth the clean fusion energy that lights the sun and stars, using this form of artificial intelligence to create a model for rapid control of plasma (the state of matter composed of free electrons and atomic nuclei, or ions) that fuels fusion reactions. Machine learning speeds modeling of experiments aimed at capturing fusion energy on Earth | DOE/Princeton Plasma Physics Laboratory That work trained a neural network to reproduce neutral beam heating calculations for the NSTX-U tokamak fast enough for real-time use. Since then, the field has moved from fast emulators to controllers that run real experiments.


Machine learning can help bring to Earth the clean fusion energy that lights the sun and stars. It does not replace plasma physics or engineering; it supplies the speed, pattern recognition, and optimization that let physicists simulate, read, and steer a plasma on the timescales the plasma itself sets.


Mechanisms

AI in fusion is best understood as a layered pipeline. Each layer can be used on its own, but the most ambitious recent experiments chain all of them together.

1. Data and simulation foundation

Learning starts with data. Tokamaks archive every pulse: magnetic probe signals, interferometer densities, Thomson scattering and electron cyclotron emission (ECE) profiles, actuator commands, and operator logs. Multi-machine databases from DIII-D, JET, KSTAR, TCV, and other devices give ML models thousands of labeled examples of disruptions, tearing modes, and edge-localized modes (ELMs).

Simulation adds breadth where experiments are thin. LLNL's ignition prediction model drew on more than 150,000 high-fidelity radiation-hydrodynamics simulations run on the Sierra supercomputer, combined with years of NIF shot data. Tokamak surrogate models are trained on large ensembles of turbulence, equilibrium, and heating calculations.

2. Surrogate models

A surrogate is a neural network trained to reproduce the output of a slow physics code. Once trained, it runs orders of magnitude faster, which makes three things possible: real-time use inside a control system, millions of what-if runs for scenario design, and gradient-based optimization.

Examples include neural network emulators of quasilinear turbulent transport models, PPPL's neutral beam network for NSTX-U, and the EFIT (equilibrium), CAKE (kinetic profile), and ION ORB (ion heat load) surrogates inside the DIII-D digital twin. Google DeepMind's open-source TORAX simulator is designed to plug ML surrogates directly into a full plasma transport calculation.

3. State estimation and virtual diagnostics

Controllers need to know the plasma's current state, but many key quantities (current profile, pressure profile, magnetic island width) are never measured directly. ML models reconstruct them from what is measured. DIII-D uses neural network kinetic equilibrium reconstruction in real time, and MIT and EPFL have demonstrated an inverse Grad-Shafranov neural network that supplies setpoints to TCV's conventional magnetic controllers. The Diag2Diag model (2025) goes further, generating synthetic super-resolution measurements from correlated signals so researchers can see events that slow diagnostics miss.

4. Forecasting instabilities and outcomes

Prediction models estimate the probability of a bad event in the near future. Princeton's Fusion Recurrent Neural Network (FRNN, published 2019) learned to forecast disruptions and showed it could transfer between DIII-D and JET. Later models forecast tearing modes ("tearability") and ELM bursts hundreds of milliseconds ahead. In inertial fusion, Bayesian ML models forecast the full probability distribution of a shot's yield rather than a single number.

5. Learned control and optimization

Decisions come from several families of algorithms:

  • Reinforcement learning (RL) in a physics simulator. DeepMind and EPFL trained a policy entirely in a free-boundary equilibrium simulator of TCV, then deployed it directly on the hardware to command all 19 magnetic control coils.
  • RL on a learned dynamics model. When no simulator is fast or accurate enough, researchers first train a neural network to predict how the plasma responds to actuators, then train the RL agent inside that model. Princeton used this approach for tearing mode avoidance on DIII-D; MIT used a physics-structured version for safe plasma shutdowns on TCV.
  • Search and optimization over scenarios. Differentiable simulators such as TORAX can be paired with RL or evolutionary search (Google DeepMind has named AlphaEvolve as one option) to explore huge numbers of candidate operating scenarios before a machine runs.
  • Human-in-the-loop optimization. TAE and Google's 2017 "Optometrist algorithm" proposed new machine settings and let a human expert choose between outcomes, which dramatically cut the number of experiments needed to tune a complex device.

6. Integration, safety layers, and the human role

Real control systems need more than a clever model. PPPL's PACMAN framework (2026) shows the emerging architecture: gather all real-time measurements, check them for errors, let multiple independent ML models read what they need and predict the plasma's behavior, let controllers turn predictions into commands, and finally pass everything through an arbitration stage that resolves conflicts and enforces hard hardware limits. Low-level loops often stay classical. Humans still set the objectives and tune the controllers between experiments.

How an AI-assisted pulse runs

  1. Before the pulse: physicists design a scenario using fast simulators or a digital twin, sometimes with an AI search over thousands of candidate trajectories.
  2. During the pulse: sensors stream data at kilohertz rates. A state estimator reconstructs the plasma, predictors flag approaching instabilities, and a policy or controller adjusts coils, heating beams, gas valves, and microwave mirrors. DeepMind's TCV policy ran at 10 kHz; PACMAN's full multi-model loop on DIII-D typically completes in about 20 milliseconds and repeats continuously.
  3. After the pulse: new data feeds back into training. PPPL reported that once its framework was in place, adding a new model took days rather than months, so teams can retrain and redeploy from week to week.

Dependencies and remaining uncertainty

Every layer depends on conditions that are not guaranteed:

  • Simulator fidelity. RL policies trained in simulation inherit the simulator's blind spots (the "sim-to-real gap").
  • Data coverage. Models trained on one machine's operating space may fail outside it.
  • Actuator authority and latency. A perfect forecast is useless if the heating systems or coils cannot respond in time.
  • Real-time computing. Controllers must run on deterministic hardware inside the plasma control system.

The largest open question is extrapolation. No device has yet operated a power-plant-scale burning plasma, where heating from fusion alpha particles dominates. Models trained on today's machines have no direct examples of that regime.

Where AI is applied in fusion research
Problem Typical AI technique Example
Plasma shape and position control Deep RL trained in a physics simulator DeepMind and EPFL on TCV (2022)
Instability avoidance RL trained on a learned dynamics model Tearing mode avoidance on DIII-D (2024)
Disruption forecasting Recurrent and convolutional neural networks FRNN on DIII-D and JET (2019)
Safe shutdown (rampdown) design Physics-structured neural state-space model plus RL MIT and EPFL on TCV (2025)
Missing or slow diagnostics Multimodal super-resolution Diag2Diag on DIII-D (2025)
Ignition and yield prediction Bayesian inference, deep learning, transfer learning LLNL cognitive simulation at NIF (2022 shot, published 2025)
Integrated scenario design Differentiable simulation plus RL or evolutionary search TORAX for SPARC (2024 to present)
Facility-wide planning AI surrogate digital twins DIII-D twin with NVIDIA; Thea Energy Helios twin (2026)
Device tuning Human-in-the-loop optimization TAE and Google Optometrist algorithm (2017)

The Physics

Fusion reactions and the energy balance

The D-T reaction produces a helium nucleus (an alpha particle carrying 3.5 MeV) and a neutron (14.1 MeV), for 17.6 MeV in total. The neutron escapes the magnetic field and carries most of the energy to the surrounding blanket, where it becomes heat and, in a power plant, breeds new tritium from lithium. The alpha particle stays in the plasma and heats it.

Whether a plasma produces net energy depends on the Lawson criterion, usually expressed as the triple product of density, temperature, and energy confinement time:

n × T × τE ≳ 3 × 1021 keV·s/m3 (for D-T ignition, near an optimum ion temperature of roughly 15 keV)

One keV corresponds to about 11.6 million kelvin, so 150 million °C is roughly 13 keV. Progress is tracked with gain factors:

  • Q (magnetic fusion): fusion power divided by external heating power. Q > 1 is scientific breakeven; ITER targets Q = 10; SPARC targets Q > 1 with design studies suggesting much higher values are possible.
  • Target gain G (inertial fusion): fusion energy out divided by laser energy delivered to the target. NIF's April 2025 record shot reached G ≈ 4.13.

Neither number accounts for the electricity needed to run magnets, lasers, and cooling systems, which is why "scientific" breakeven and "engineering" breakeven are very different milestones.

Magnetic confinement, equilibrium, and stability limits

Charged particles spiral around magnetic field lines, so a strong field slows their escape. In a tokamak, external coils create a toroidal field and a large current flowing in the plasma creates a poloidal field; together they form nested, twisted magnetic surfaces. Force balance (∇p = J × B) leads to the Grad-Shafranov equation, which describes the plasma's equilibrium shape. Reconstructing that equilibrium from magnetic sensor readings is an inverse problem that must be solved continuously during a pulse, which is one reason ML has been so useful.

Several dimensionless limits define the safe operating space:

  • Plasma beta (β = 2μ0⟨p⟩ / B2) compares plasma pressure to magnetic pressure. At fixed temperature, fusion power density scales roughly as β2B4. That steep dependence on field strength is why CFS built SPARC around high-temperature superconducting magnets (about 12.2 T on axis), allowing a much smaller device.
  • Safety factor q measures how tightly field lines twist. Low edge q improves performance but raises the risk of kink and tearing instabilities.
  • The Greenwald density limit (nG in 1020 m-3 = Ip in MA / (π a2), where a is the minor radius) is an empirical ceiling above which tokamak plasmas tend to disrupt. Recent EAST results (see below) suggest it can be exceeded under the right wall conditions.

Instabilities and disruptions: the control problem

The physics that makes AI control attractive is a set of instabilities that grow fast and interact:

  • Tearing modes occur when the plasma current profile lets magnetic field lines break and reconnect into "magnetic islands." The classical stability parameter Δ′ predicts some of this behavior, but real tearing onset depends on many coupled profiles, which is why data-driven forecasts help.
  • Edge-localized modes (ELMs) are repeated bursts from the steep pressure "pedestal" at the plasma edge. Large ELMs can erode wall components in a reactor.
  • Alfvén eigenmodes are waves driven by fast particles (such as beam ions or, in a reactor, alpha particles) that can eject those particles before they heat the plasma.
  • Vertical instability affects elongated plasmas, which naturally tend to drift up or down within milliseconds unless the coils actively hold them.
  • Disruptions end a pulse abruptly. The plasma's thermal energy is lost in about a millisecond (the thermal quench), the current then decays (the current quench), large electromagnetic forces hit the vessel, and runaway electrons can form. For ITER and future power plants, disruptions are among the most serious operational risks.

These events span a huge range of timescales: Alfvénic dynamics in microseconds, MHD instability growth in milliseconds, energy confinement in tenths of a second to seconds, and current diffusion over seconds or longer. No single first-principles code covers all of them fast enough for control, which is the gap ML fills.

Transport, turbulence, and why surrogates are needed

Most heat leaks out of a tokamak through small-scale turbulence driven by temperature and density gradients (ion temperature gradient, trapped electron, and electron temperature gradient modes). Gyrokinetic codes that simulate this turbulence directly are extremely expensive. Quasilinear models (such as QuaLiKiz and TGLF) are cheaper but still too slow to call millions of times inside an integrated simulation. Neural network surrogates of those quasilinear models evaluate almost instantly, making it practical to simulate an entire discharge quickly enough for optimization loops and, in some cases, real-time control.

Heat exhaust adds an engineering constraint. The power leaving the plasma concentrates on narrow strips of the divertor, the component that handles exhaust. One mitigation is to sweep the magnetic strike point back and forth to spread the heat, which is one of the first RL targets in the DeepMind and CFS collaboration.

Inertial confinement: sensitivity and uncertainty

At NIF, 192 laser beams heat a small gold cylinder (a hohlraum), which converts the laser light into X-rays. The X-rays ablate the outer surface of a fuel capsule, driving the fuel inward at very high speed. Small imperfections in the capsule, asymmetries in the drive, and hydrodynamic instabilities (notably Rayleigh-Taylor) can sharply reduce yield. Because outcomes are so sensitive to details that cannot be fully measured in advance, the useful question is not "what will the yield be?" but "what is the probability distribution of yields?" Bayesian ML methods that calibrate simulations against past shots are designed to answer exactly that question.

Learning as an engineering tool

Three mathematical ideas recur:

  • Reinforcement learning frames control as a Markov decision process. A policy maps observations to actions and is trained to maximize a reward that encodes goals (track a target shape, maximize pressure) and penalties (approach to instability, violation of coil current limits). DeepMind's TCV work used an actor-critic design in which a small actor network runs fast enough for real-time control while a larger critic network, used only during training, evaluates the actor's choices.
  • Bayesian inference and transfer learning combine many cheap but biased simulations with a few expensive, accurate experiments, producing calibrated predictions with uncertainty bands.
  • Multimodal learning exploits the fact that different diagnostics observe the same physical system. A fast but indirect signal (such as ECE, which tracks electron temperature) can help infer what a slower, more complete diagnostic (such as Thomson scattering, often limited to a few hundred hertz) would have seen between its samples.

Why This Matters

Scientific importance

AI is starting to reveal physics rather than only accelerate known calculations. Diag2Diag's super-resolution reconstructions captured fast edge-profile changes during ELM suppression that existing diagnostics could not resolve, supporting the hypothesis that magnetic islands play a role in suppressing ELMs. Physics-structured models, such as MIT's neural state-space model for TCV, show that combining known equations with learned corrections can produce reliable predictions from a few hundred pulses rather than thousands. Researchers are also working on interpretable ML metrics that tell operators how close a plasma is to a stability boundary, not just whether an alarm should sound.

Engineering and operations

A commercial fusion plant has to run near its performance limits, for long periods, with very few disruptions. AI contributes in several concrete ways:

  • Wider safe operating space. Controllers that anticipate instabilities can hold a plasma at higher pressure than conservative preprogrammed trajectories allow.
  • Fewer wasted experiments. Predict-first design, digital twins, and probabilistic shot forecasts let teams spend scarce machine time on the most promising experiments.
  • Faster commissioning. CFS uses TORAX to run large numbers of virtual SPARC experiments before first plasma, so operators start with tested plans.
  • Leaner diagnostics. Virtual diagnostics can identify the minimum set of sensors a power plant needs and fill gaps when a sensor fails.

Commercial and energy-system implications

The relationship between AI and fusion runs in both directions. AI developers need large amounts of firm, carbon-free power for data centers, and several of them have become fusion customers and investors. Google agreed to buy 200 MW from CFS's first ARC plant in Virginia (planned for the early 2030s); Microsoft signed a 2023 agreement with Helion for fusion electricity targeted for 2028; and Google has backed TAE for more than a decade. Private fusion funding reached about $15 billion by September 2025, according to a report by the EU body Fusion for Energy. Grid operators and utilities are starting to treat fusion as a potential long-term resource, though no fusion plant has yet delivered electricity.

Policy and research infrastructure

The U.S. Department of Energy now treats AI as a core part of its fusion strategy. Its October 2025 Fusion Science and Technology Roadmap calls for an "AI-Fusion Digital Convergence" platform that links fusion facilities, national laboratories, supercomputers, and private companies. The Genesis Mission, launched in November 2025, lists fusion among its priority areas and funded its first fusion-related AI projects in July 2026.

Limitations

AI does not solve fusion's hardest non-plasma problems: materials that survive years of 14 MeV neutron damage, tritium breeding and handling, heat exhaust at power-plant scale, plant reliability, and licensing. Most AI control demonstrations so far ran on small and medium research tokamaks, for limited experiment campaigns, under close expert supervision. Learned controllers also raise verification questions: regulators and operators will need ways to certify that a neural network will behave safely in conditions it has never seen.

What would change if AI-for-fusion succeeds

If learned control and simulation prove reliable at reactor scale, machines could run closer to their limits, reach target performance with fewer commissioning shots, and avoid more disruptions, all of which improve plant availability and the cost of electricity. A compact machine with excellent control could match the performance of a larger, more conservatively operated one. If AI methods do not generalize to burning plasmas, they will still speed up research, but plants may need larger safety margins and more conventional, physics-based control.

Research Landscape

Recent Developments

Instability avoidance with reinforcement learning on DIII-D (2024)

What was tested: Princeton and PPPL researchers, led by Jaemin Seo and Egemen Kolemen, asked whether an RL controller could keep a high-performance plasma away from tearing instabilities, a leading cause of disruptions, especially in ITER-like scenarios.

Method: They first trained a multimodal neural network on past DIII-D data to estimate "tearability," the likelihood of a tearing mode in the near future, from diagnostic and actuator signals. They then used that model as the training environment for an RL agent that adjusted beam power and plasma shape.

Findings: In DIII-D experiments, the controller kept tearing likelihood below a set threshold while maintaining high-confinement (H-mode) performance, even under unfavorable conditions of low safety factor and low torque. The predictive model could anticipate tearing up to about 300 milliseconds ahead. Results were published in Nature in February 2024.

Limits: The demonstration was on one machine and a limited set of scenarios. The authors stated that more data is needed to show the controller works across different conditions and devices.

PACMAN: integrating many AI models into one control loop (2026)

What was tested: Whether multiple independent ML models could run together inside a tokamak's real-time control system, rather than as one-off demonstrations.

Method: PPPL and Princeton developed PACMAN (Prediction And Control using MAchiNe learning), a modular framework with distinct stages for data gathering, validation, model prediction, control, and safety arbitration. The team, including co-lead authors Hiro Farre Kaga and Andy Rothstein, tested it in five DIII-D experiments, published in Nuclear Fusion in September 2026.

Findings: During those experiments, PACMAN let an RL policy take full control of the heating systems; predicted ELM bursts; detected and controlled fast-particle-driven waves; steered plasma density and rotation to targets; and predicted a tearing mode about 200 milliseconds ahead so the plasma could avoid it. It also coordinated all six of DIII-D's gyrotrons (microwave heating sources), adjusting mirror angles and power together, a task for which no conventional algorithm existed. The full loop typically runs in about 20 milliseconds.

What it does and does not establish: PACMAN shows that AI control can become reusable infrastructure: the first model took months to install, the second took days. It does not yet show long-term reliability, transfer to other machines, or performance in reactor conditions. Safety limits and human-set objectives remain in charge.

Physics-structured learning for safe rampdowns on TCV (2025)

Shutting a tokamak down (ramping down the plasma current) is surprisingly risky and hard to simulate. MIT researchers led by Allen Wang, working with EPFL, built a neural state-space model that embeds known physics structure and learns corrections from data. Trained on a few hundred TCV pulses, it can simulate on the order of ten thousand rampdown trajectories per second on a single GPU, fast enough to train an RL agent. RL-designed rampdowns produced statistically significant improvements in TCV experiments. In a predict-first test, the team raised plasma current about 20 percent above the training range (from 140 kA to 170 kA) and generated the trajectory entirely from model predictions before the experiment. The work, published in Nature Communications in October 2025, was motivated by preparation for SPARC operations. The extrapolation demonstrated was modest, and the authors framed it as evidence of small, reliable extrapolations rather than general transfer.

AI-boosted diagnostics: Diag2Diag (2025)

An international team from Princeton, PPPL, Chung-Ang University, Columbia University, and Seoul National University, led by Azarakhsh Jalalvand, trained Diag2Diag on DIII-D data to learn correlations among diagnostics. The model generated synthetic, super-resolution Thomson scattering data from faster signals (ECE, interferometry, and magnetics), capturing millisecond-scale ELM dynamics that the real Thomson system samples too slowly to see. Published in Nature Communications in 2025, it is a proof of concept. Its value for reactors lies in two places: recovering data when a sensor fails and helping identify the minimum diagnostic set a power plant will need. Synthetic measurements are still model outputs, so their reliability outside the training distribution needs independent checks.

Predicting ignition at NIF (published 2025)

LLNL's cognitive simulation team combined a large simulation database, Bayesian post-processing, and transfer learning into a generative model of implosion performance. Days before the December 5, 2022 shot, the model estimated a 74 percent probability of exceeding breakeven, far higher than for earlier designs, and the measured yield (3.15 MJ from 2.05 MJ of laser energy) fell within its predicted range. The detailed method was published in Science in 2025, and LLNL reports testing the model on many later D-T experiments. NIF performance has continued to rise: 5.2 MJ in February 2024 and a record 8.6 MJ in April 2025 with a target gain of about 4.13. LLNL is now developing an Enhanced Yield Capability that would raise laser energy from about 2.2 MJ to 2.6 MJ. A prediction model that is right about a probability does not guarantee any single shot, and NIF remains a single-shot facility, far from the repetition rate an inertial fusion power plant would need.

Differentiable simulation and the DeepMind-CFS partnership (2025 to 2026)

After its 2022 TCV result, Google DeepMind released TORAX in 2024: an open-source, differentiable tokamak transport simulator written in JAX that runs on CPUs and GPUs and integrates ML surrogates. A 2024 follow-up paper on TCV also improved the precision and training efficiency of RL magnetic control. In October 2025, DeepMind and CFS formalized a research partnership focused on three areas:

  • building a fast, accurate, differentiable plasma simulation for SPARC;
  • searching for the most efficient and robust path to maximum fusion energy, using TORAX with RL or evolutionary search such as AlphaEvolve;
  • developing RL-based real-time control, starting with strategies for sweeping exhaust heat across SPARC's divertor tiles and quickly tuning conventional controllers for specific pulses.

CFS staff have said TORAX saves substantial setup time for SPARC simulations. SPARC itself (a compact, high-field tokamak with a plasma current of about 8.7 MA) was roughly 80 percent complete by mid-2026, with the first of 18 toroidal field magnets installed in January 2026 and first plasma and Q > 1 targeted for 2027. Validating TORAX against past tokamak data and high-fidelity simulations is a stated priority, since SPARC's first real data will test every pre-built model.

Digital twins

  • DIII-D. General Atomics and NVIDIA have built an interactive digital twin of DIII-D in NVIDIA Omniverse, with AI surrogates (EFIT, CAKE, ION ORB) trained on decades of data using the Polaris supercomputer at Argonne and Perlmutter at NERSC. The aim is to cut some simulation turnaround from weeks to seconds so the facility's roughly 700 collaborators can test ideas between shots. Argonne reports that the twin is being validated against the operating facility and was demonstrated at the Confab26 conference.
  • Stellarators. In June 2026, Thea Energy announced a digital twin of its Helios stellarator power plant concept with NVIDIA, Synopsys, Argonne (neutronics and blanket design), and PPPL, as part of the Genesis Mission.
  • NSTX-U and engineering tools. PPPL is building a digital twin of its NSTX-U spherical tokamak, and in 2025 PPPL, CFS, and Oak Ridge described an ML tool that finds "shadowed" regions on plasma-facing components much faster than conventional heat-load calculations.

A twin is only as good as its validation. Most published twin work so far describes architecture and early demonstrations rather than peer-reviewed accuracy benchmarks.

National programs: the DOE roadmap and the Genesis Mission

DOE's Fusion Science and Technology Roadmap (October 16, 2025) set a "Build, Innovate, Grow" strategy targeting commercial fusion by the mid-2030s, with AI-fusion convergence threaded through all program elements. The Genesis Mission, launched by executive order on November 24, 2025, named fusion among its energy priorities. In February 2026 DOE published 26 national science and technology challenges, including an AI-Fusion Digital Convergence Platform. In July 2026 DOE selected 278 Genesis Mission projects across many fields. Fusion-related selections include a PPPL-led project to build an AI system that autonomously operates high-power gyrotrons for electron cyclotron heating, a University of Rochester Laboratory for Laser Energetics project on agentic AI for inertial fusion implosion design, and PPPL work on the STELLAR-AI computing platform and a seed "Fusion-FM" foundation model for magnetic confinement. Supporters see the program as a way to sharply increase research productivity; some critics have questioned how it fits alongside reductions in other federal science budgets.

Private sector developments

  • TAE Technologies and Google. Google engineers worked on-site at TAE to co-develop plasma reconstruction algorithms. TAE credits insights from that work with its 2023 decision to simplify its machine design, leading to "Norm," which formed and stabilized plasma above 70 million °C using neutral beams alone. In June 2025 TAE raised more than $150 million with Google and Chevron participating. In December 2025 TAE agreed to an all-stock merger with Trump Media & Technology Group valued at more than $6 billion; as of mid-2026 the companies were targeting a close in the fourth quarter of 2026.
  • Helion. In February 2026 Helion announced that its Polaris prototype reached plasma temperatures of 150 million °C and produced measurable D-T fusion, which it described as a first for a privately developed machine. Construction of its Orion plant in Malaga, Washington, began in July 2025 to serve its 2028 Microsoft agreement. Many observers consider that date very aggressive.
  • Microsoft. Microsoft Research held its first Fusion Summit in March 2025, signed a memorandum of understanding with PPPL covering plasma control, digital twins, materials, and experiment optimization, and is working with ITER on using AI in the project's design and operations.

Plasma performance milestones that raise the stakes for control

  • Long pulses. China's EAST sustained a high-confinement plasma for 1,066 seconds in January 2025. France's WEST then held a plasma for 1,337 seconds (about 22 minutes) on February 12, 2025, with 2.6 GJ of injected and extracted energy.
  • Stellarator performance. Wendelstein 7-X set a world record for the triple product in long discharges on May 22, 2025, sustaining it for 43 seconds while injecting about 90 frozen fuel pellets.
  • Density. In January 2026 the EAST team reported start-up plasmas at line-averaged densities of 1.3 to 1.65 times the Greenwald limit, using electron cyclotron heating and higher prefill gas pressure, in line with plasma-wall self-organization theory. The result applied to start-up plasmas; testing in high-confinement mode is the next step.
  • ITER schedule. ITER's 2024 baseline replaced the original 2025 first-plasma target with a Start of Research Operations in 2034 (with full 15 MA operation and a tungsten first wall) and D-T operation planned for 2039. ITER's long timeline is one reason AI-ready simulation and control tools are being developed on smaller machines first.

Long pulses, high density, and high performance all make control harder: there are more chances for an instability and less tolerance for a disruption. These records define the operating regimes that AI controllers must eventually handle.

Timeline of Key Results

Year Group Device AI approach Result
2017 TAE Technologies and Google C-2U Optometrist algorithm (human-in-the-loop optimization) Sustained net plasma heating; far fewer tuning experiments
2018 JET contributors JET Deep learning (CNNs, RNNs) Plasma tomography and disruption prediction
2019 Princeton / PPPL DIII-D, JET Fusion Recurrent Neural Network Cross-machine disruption prediction
2019 PPPL NSTX-U Neural network surrogate Real-time-capable neutral beam modeling
2022 Google DeepMind and EPFL TCV Deep RL; one policy for all 19 coils Diverse shapes, including two simultaneous "droplet" plasmas
2022 (published 2025) LLNL NIF Bayesian deep learning (cognitive simulation) 74% pre-shot ignition probability; yield within prediction
2024 Princeton / PPPL / DIII-D DIII-D RL on learned tearability model Tearing avoidance with H-mode maintained
2024 Google DeepMind (software) TORAX differentiable simulator Open-source JAX transport code with ML surrogates
2025 Princeton-led international team DIII-D Multimodal super-resolution (Diag2Diag) Synthetic high-speed Thomson data; ELM physics revealed
2025 MIT and EPFL TCV Neural state-space model plus RL Robust rampdowns; predict-first experiment at +20% current
2025 Google DeepMind and CFS SPARC (preparation) TORAX with RL and evolutionary search Research partnership for scenario design and control
2026 PPPL / Princeton DIII-D PACMAN modular multi-model control Five experiments; ~20 ms loop; tearing predicted ~200 ms ahead

Open Questions and Limitations

  • Extrapolation to burning plasmas. Alpha-particle self-heating changes plasma dynamics, and no existing machine provides training data in that regime. ITER and SPARC will be the first real tests.
  • Cross-machine transfer. FRNN showed promising transfer between DIII-D and JET, but most control results remain single-machine. Each new device has different geometry, actuators, and diagnostics.
  • Data scarcity and sharing. Fusion data is small by ML standards and spread across institutions. Physics-informed models, foundation models such as the proposed Fusion-FM, and transatlantic computing links (such as the plan PPPL and the UK Atomic Energy Authority announced in 2026) aim to help, but none has yet been proven at scale.
  • Verification and trust. Learned controllers need formal safety envelopes, clear failure modes, and interpretability before operators and regulators will accept them in a power plant. Current practice keeps hard limits and human oversight in place.
  • Simulator bias. Optimizing against a simulator can exploit its errors. DeepMind's early TCV work found that an unconstrained agent used coils in ways the engineers had not intended, which had to be removed from its action space.
  • Hype versus evidence. Press coverage sometimes claims AI will cut years off fusion timelines. Such estimates are not derived from peer-reviewed analysis, and fusion's non-plasma challenges (materials, tritium, plant engineering) are largely outside what today's AI control addresses.
  • Competition for capital. Fusion investment and AI investment are intertwined. AI companies are major fusion customers, but some fusion industry observers note that AI's dominance of venture funding can also draw money away from fusion developers.

Resources

Additional Viewing

Load video
YouTube

How AI Predicted Fusion Ignition Before It Happened (Lawrence Livermore National Laboratory, Data Science Institute seminar, February 2023). LLNL physicists Kelli Humbird and J. Luc Peterson explain the December 2022 NIF ignition shot and how the cognitive simulation workflow combined simulations, experimental data, and machine learning to forecast that ignition was the most likely outcome. It is the best companion to the "Predicting ignition at NIF" section above.

Load video
YouTube

The Path to Safe Fusion Energy through Deep Learning (San Diego Supercomputer Center, May 2019). William Tang of Princeton describes the neural network his team built to predict plasma disruptions, the work published in Nature as the Fusion Recurrent Neural Network. The talk explains why disruptions are so damaging and how training at supercomputer scale enabled accurate, cross-machine forecasts.

Load video
YouTube

At the intersection of TensorFlow and nuclear physics (TensorFlow Meets, June 2018). Google software engineer Ian Langmore talks with host Laurence Moroney about using TensorFlow to reconstruct plasma measurements in TAE Technologies' plasma generator. It is useful background on the Google and TAE collaboration that later informed TAE's simplified "Norm" design.

Load video
YouTube

Can AI help crack the code of fusion power? (Verge Science). This companion to The Verge's 2019 article visits TAE Technologies to show how Google's machine-assisted optimization helped tune a complex fusion experiment. It is an accessible introduction to why AI appealed to fusion researchers before the recent wave of control results.

Load video
YouTube

Retained from the earlier version of this page's video collection on AI and fusion research.

Ongoing talks:

References and Further Reading

Foundational studies

Magnetic confinement control and diagnostics (2024 to 2026)

Inertial confinement fusion

Simulation, digital twins, and industry partnerships

Policy and national programs

Plasma performance and facility milestones

Retrieved from "https://primo.ai/w/index.php?title=Nuclear_Fusion&oldid=49847"