DEV Community

dosanko_tousan
dosanko_tousan

Posted on

AI and Nuclear Fusion Vol.5: AI-Accelerated Fusion — Compressing the Timeline

#nuclear #ai #machinelearning #plasma

AI and Nuclear Fusion Vol.5: AI-Accelerated Fusion — Compressing the Timeline

Series: "Thinking Seriously About Nuclear Fusion with AI"
Volume 5 of 10 | Target: Policy, Investment, and Engineering Decision-Makers
Author: Dosanko Tousan | AI Partner: Claude (Anthropic)
License: MIT

Document Classification

Item Detail
Purpose Assess where AI/ML can realistically compress the fusion engineering timeline and where it cannot; quantify the gap between AI hype and demonstrated impact in fusion research
Audience Fusion program managers, AI/ML researchers considering fusion applications, policy advisors evaluating AI-fusion investment, venture capitalists
Prerequisites Vol.1–4 (physics, ignition, tritium, materials). AI/ML fundamentals helpful but not required — all methods explained from first principles.
Scope Disruption prediction → Real-time plasma control → Material discovery → Reactor design optimization → Stellarator coil optimization → Limitations and honest assessment
Deliverables (1) Disruption prediction demonstration code, (2) Material screening surrogate model, (3) Timeline compression assessment, (4) Decision Matrix for AI-fusion investment

Table of Contents

Part I: The Case for AI in Fusion

Part II: Plasma Control and Disruption Avoidance

Part III: Accelerating Materials and Design

Part IV: The Honest Assessment

§1. Executive Summary

AI can compress the fusion timeline. It cannot replace the experiments we have not done.

Volumes 3 and 4 of this series established two engineering crises: the tritium fuel supply runs out around 2045–2050, and structural materials have been validated to only one-fifth of their required lifetime. Both crises are governed by timelines that physics breakthroughs alone cannot accelerate — they require engineering validation, industrial deployment, and irradiation data that accumulate at the speed of neutrons, not the speed of thought.

This volume asks: Can artificial intelligence compress those timelines?

The answer is nuanced. AI is already delivering measurable impact in specific fusion domains:

Where AI is working (demonstrated, published, reproduced):

  • Disruption prediction in tokamaks: ML models achieve >95% detection accuracy with >30 ms warning time on JET, DIII-D, and EAST — sufficient for mitigation system activation
  • Plasma shape and position control: DeepMind's reinforcement learning controller on TCV (2022) demonstrated simultaneous multi-objective plasma control — a genuine breakthrough in real-time AI application
  • Stellarator coil optimization: AI-driven optimization has reduced the design space from astronomically large to computationally tractable, enabling configurations like QUASAR and next-generation stellarator concepts
  • Surrogate models for neutronics: Neural networks trained on MCNP/TRIPOLI outputs can replace hours-long Monte Carlo simulations with millisecond inference, enabling design space exploration that was previously impossible

Where AI is promising but unproven:

  • Material property prediction for fusion conditions (no training data above 20 dpa at correct He/dpa)
  • Real-time burn control in ignited plasmas (no ignited plasma exists yet to train on)
  • Tritium inventory management in breeding blankets (no operating blanket exists)
  • Disruption prediction on ITER (no ITER plasma data exists; transfer learning from smaller machines is unvalidated)

Where AI cannot help:

  • Producing the 14 MeV neutron irradiation data that does not yet exist
  • Replacing material irradiation experiments with predictions (the physics of damage at >20 dpa with fusion He/dpa is unknown, not merely uncalculated)
  • Eliminating the tritium cliff through computation
  • Substituting for engineering integration tests of systems that have never been built

Key finding: AI's greatest value in fusion is not in replacing experiments but in reducing the number of experiments needed and extracting maximum information from each experiment performed. Given that the fusion program's binding constraint is time (the tritium cliff, the DONES data delivery timeline, the DEMO construction schedule), any technology that compresses the experimental-validation loop provides existential leverage.

Quantitative timeline compression estimate:

Domain Without AI With AI (estimated) Compression
Disruption avoidance system development 10–15 years 5–8 years 40–50%
Stellarator coil optimization 20+ years 5–10 years 50–75%
Material screening (candidate identification) 10–15 years 3–5 years 60–70%
Material qualification (irradiation testing) 15–20 years 12–18 years 10–20%
Breeding blanket TBR optimization 5–10 years 2–4 years 50–60%
Reactor design space exploration 10–15 years 3–5 years 60–70%
Overall program timeline 25–30 years 18–23 years ~25%

The overall compression is limited by the rate-limiting step: irradiation testing and engineering integration, which AI can optimize but not eliminate. A 25% reduction in the overall fusion timeline — roughly 5–7 years — is significant enough to determine whether DEMO operates within the tritium supply window or misses it entirely.

Part I: The Case for AI in Fusion

§2. Why Fusion Needs AI — The Complexity Argument

Fusion plasma is the most complex physical system that humans attempt to control in real time. To appreciate why AI is not merely helpful but arguably necessary, consider the dimensionality of the problem:

The plasma state space:

A tokamak plasma is described by:

  • Electron density profile $n_e(r, \theta, \phi, t)$: ~10²⁰ m⁻³, varying across the plasma cross-section
  • Ion density profiles $n_i(r, t)$ for each species (D, T, He, impurities)
  • Electron temperature profile $T_e(r, t)$: 0.1–15 keV
  • Ion temperature profile $T_i(r, t)$: similar range
  • Current density profile $j(r, t)$: determines magnetic geometry
  • Magnetic field profiles $B_\phi(r)$, $B_\theta(r)$: 2–12 T
  • Rotation profiles (toroidal and poloidal)
  • Impurity concentration profiles (C, O, W, Be, etc.)
  • Neutral beam deposition profile
  • RF heating deposition profile
  • Edge pedestal structure (H-mode)
  • ELM characteristics (frequency, amplitude, type)

Each profile is effectively a continuous function of radius and time. Discretized on a typical numerical grid (100 radial × 64 poloidal × 1000 time steps), the state space for a single discharge is ~10⁷ dimensions.

The control input space:

The plasma physicist has access to:

  • Ohmic transformer (coil current → plasma current)
  • Poloidal field coils (typically 6–12 independent circuits → plasma shape and position)
  • Neutral beam injectors (4–8 beamlines → heating, current drive, rotation)
  • RF systems (ECRH, ICRH, LHCD → localized heating and current drive)
  • Gas injection (multiple species, multiple locations → density control)
  • Pellet injection (fuel and impurity pellets → deep fueling, ELM pacing)
  • Divertor pumping (→ particle exhaust)
  • Error field correction coils (→ stability)

Each actuator has constraints (maximum power, ramp rate, time delay), and many interact nonlinearly with the plasma state.

The optimization objective:

The control system must simultaneously:

  1. Maintain plasma position within ±1 cm (to avoid wall contact)
  2. Maintain plasma current at the target value (±1%)
  3. Achieve and sustain target density and temperature profiles
  4. Avoid disruptions (catastrophic plasma termination)
  5. Manage impurity accumulation
  6. Control ELM behavior (if in H-mode)
  7. Maximize fusion power while respecting operational limits
  8. Minimize divertor heat flux

This is a high-dimensional, nonlinear, multi-objective, real-time control problem with partial observability (diagnostic measurements are noisy and incomplete), actuator constraints, and catastrophic failure modes (disruptions). The time scale for plasma instabilities can be as short as milliseconds, while the thermal and current profile evolution occurs over seconds.

Traditional control approaches — primarily PID controllers and linear MHD models — handle the well-understood aspects (vertical position, plasma current) but struggle with:

  • Nonlinear coupling between profiles
  • Proximity to stability boundaries
  • Adaptation to changing wall conditions (over weeks/months)
  • Multi-objective optimization under constraints

This is precisely the problem class where modern AI/ML excels: learning complex nonlinear mappings from high-dimensional data, optimizing multi-objective functions, and making real-time decisions under uncertainty.

§3. The Landscape: Who Is Doing What

The fusion-AI landscape as of 2025 spans academic research, national laboratories, and private companies:

National laboratories and large experiments:

Institution Experiment AI/ML Focus
PPPL (USA) NSTX-U Disruption prediction, surrogate models, data-driven transport
MIT / CFS SPARC ML-assisted design, HTS magnet optimization
CCFE (UK) JET / MAST-U Disruption prediction (FREIA), real-time control
IPP (Germany) Wendelstein 7-X Stellarator optimization, equilibrium reconstruction
ITER Organization ITER Integrated modeling, disruption mitigation system
CEA (France) WEST Real-time plasma control, wall condition monitoring
ASIPP (China) EAST Long-pulse control, ML-based profile prediction
QST (Japan) JT-60SA Scenario optimization, disruption prediction
KFE (Korea) KSTAR AI-based disruption prediction, plasma imaging

Private companies with significant AI/ML integration:

Company Fusion concept AI/ML role
Commonwealth Fusion Systems HTS tokamak (SPARC/ARC) Magnet design, plasma scenario optimization
TAE Technologies FRC (C-2W, Copernicus) Optometrist algorithm: ML-driven plasma optimization in real time; company was co-founded by ML researchers
Helion Energy FRC (pulsed, Polaris) Pulse timing optimization, target plasma conditions
DeepMind (Alphabet) N/A (research) Reinforcement learning for TCV plasma control
Google (DeepMind) N/A (research) Accelerated equilibrium solving
Microsoft N/A (partnership with Helion) Computational resources, plasma simulation acceleration
Renaissance Fusion Stellarator (HTS) AI-optimized stellarator design

Key observation: The most aggressive AI integration is in private fusion companies, where the absence of legacy control systems and the pressure to move fast create a natural testbed for ML-first approaches. TAE Technologies in particular has built its entire experimental optimization loop around machine learning, using the "optometrist algorithm" (which presents the plasma with binary choices and evolves toward optimal conditions through iterative optimization).

Part II: Plasma Control and Disruption Avoidance

§4. The Disruption Problem

A disruption is the sudden, uncontrolled termination of a tokamak plasma. It is the most dangerous operational event in a tokamak and the primary engineering design driver for ITER and DEMO first wall and vacuum vessel structural integrity.

What happens during a disruption:

  1. Thermal quench (TQ): The plasma thermal energy (up to 350 MJ in ITER) is dumped onto plasma-facing components in 1–5 ms. Peak heat fluxes can exceed 10 GW/m² — enough to melt or ablate tungsten.

  2. Current quench (CQ): The plasma current (15 MA in ITER) decays to zero in 10–50 ms. The rapidly changing current induces eddy currents and electromagnetic forces in the vacuum vessel and in-vessel components. These forces can reach thousands of tonnes — sufficient to structurally damage the machine.

  3. Runaway electrons (RE): During the current quench, a fraction of electrons can be accelerated to relativistic energies (up to ~20 MeV) by the decaying electric field. A runaway electron beam carrying 1–10 MA of current at ~20 MeV deposits its energy in a localized area, potentially melting through the first wall.

ITER disruption loads:

Parameter Value
Stored thermal energy 350 MJ
Stored magnetic energy 395 MJ
Plasma current 15 MA
Thermal quench duration 1–5 ms
Current quench duration 36–150 ms
Peak wall heat flux (TQ) >10 GW/m²
Electromagnetic forces on VV ~4,000 tonnes (vertical)
Runaway electron energy Up to 300 MJ

ITER is designed to withstand a limited number of disruptions (the design target is <10% disruption rate, with no more than a few full-energy disruptions per operational phase). A commercial power plant must have a disruption rate effectively zero — perhaps <0.1% — because each disruption causes measurable damage and requires inspection and potential repair.

Why traditional approaches are insufficient:

Disruptions are preceded by precursor events (locked modes, radiation peaking, density limit approach, etc.), but the causal chains are complex, nonlinear, and discharge-specific. A disruption can originate from:

  • Density limit exceedance
  • Vertical displacement event (VDE)
  • Locked mode instability
  • Beta limit exceedance
  • Impurity accumulation
  • Edge instabilities

Each pathway has different observable signatures, different time scales, and different optimal mitigation responses. Rule-based systems ("if locked mode amplitude > threshold, then trigger mitigation") work for well-characterized disruptions but miss novel disruption chains and edge cases. This is a classification problem with high-dimensional inputs, multiple failure modes, and catastrophic consequences for missed detections.

§5. ML for Disruption Prediction

Machine learning disruption predictors represent the most mature application of AI in fusion, with over a decade of development and deployment across multiple machines.

The classification problem:

Given a time series of diagnostic signals (plasma current, locked mode amplitude, radiated power, density, internal inductance, beta, etc.), classify each time step as:

  • Disruptive (disruption will occur within a specified warning time)
  • Non-disruptive (safe operation)

Performance metrics:

  • True positive rate (TPR): Fraction of disruptions correctly predicted
  • False positive rate (FPR): Fraction of safe discharges incorrectly flagged
  • Warning time: Time between prediction and disruption onset
  • Minimum warning time: Earliest time at which mitigation can be triggered

For a mitigation system (e.g., massive gas injection or shattered pellet injection) to prevent damage, the warning time must exceed the system activation time (~30 ms for ITER's disruption mitigation system).

State of the art (2025):

Machine Method TPR FPR Warning time Reference
JET Random Forest (FREIA) 93–97% 3–5% >30 ms de Vries et al. (2021)
DIII-D Deep learning (DPRF) 95% 5% >30 ms Rea et al. (2019)
EAST LSTM network 92–95% 5–8% >100 ms Zheng et al. (2020)
KSTAR CNN on ECE images 90–93% 5–7% >50 ms Kwon et al. (2021)
ASDEX-U Gradient boosted trees 94% 4% >20 ms Pau et al. (2019)

The transfer learning challenge:

Training a disruption predictor on Machine A and deploying it on Machine B is the central unsolved problem. Plasma behavior differs between machines due to size, magnetic field, wall material, heating systems, and operational scenarios. A model trained on JET data performs poorly on DIII-D without adaptation.

This is critical for ITER: there is no ITER disruption data to train on. The ITER disruption predictor must be built from transfer learning across multiple smaller machines, validated on early ITER hydrogen plasmas, and then deployed for D-T operations. The AI research question is whether deep learning representations are sufficiently general to capture machine-independent disruption physics.

Recent progress on cross-machine transfer:

  • Domain adaptation techniques (adversarial training, feature alignment) have shown improved transfer from JET → DIII-D
  • Physics-informed features (dimensionless parameters like β_N/β_limit, n/n_Greenwald) transfer better than raw signals
  • Foundation model approaches (pre-training on multi-machine datasets) are under active development

The ITER deployment path:

  1. Pre-plasma (now–2035): Train on JET, DIII-D, EAST, KSTAR, JT-60SA data. Develop transfer learning framework.
  2. Hydrogen/helium operations (2035–2039): Collect ITER-specific plasma data. Fine-tune models.
  3. D-D operations (2039–2042): Validate with nuclear environment. Update for new diagnostics.
  4. D-T operations (2042+): Deploy for full-energy disruption prediction. Update continuously.

The model must be deployment-ready by Phase 3 at the latest — approximately 15 years from now. This timeline is aggressive but feasible given the current pace of development.

§6. Real-Time Plasma Control with Reinforcement Learning

Classical plasma control in tokamaks uses a hierarchy of feedback loops:

  • Fast loops (μs): Vertical stability (proportional control on vertical position)
  • Medium loops (ms): Plasma current, shape, position (PID controllers)
  • Slow loops (s): Density, temperature, q-profile (model-based or empirical)

Each loop is typically designed independently, with limited coordination. The result is a control system that is robust for nominal conditions but brittle near operational boundaries — precisely where fusion reactors must operate to maximize performance.

The reinforcement learning approach:

Reinforcement learning (RL) replaces the hand-designed control hierarchy with a learned policy that maps the full observed plasma state directly to actuator commands, optimizing a multi-objective reward function that encodes all operational goals simultaneously.

Formal setup:

  • State $s_t$: Diagnostic measurements at time $t$ (density, temperature, current, shape descriptors, etc.)
  • Action $a_t$: Actuator commands (coil currents, NBI power, gas valve voltages, etc.)
  • Reward $r_t$: Weighted sum of objectives: $r = w_1 r_{\text{shape}} + w_2 r_{\text{current}} + w_3 r_{\text{stability}} + ...$
  • Policy $\pi_\theta(a_t | s_t)$: Neural network mapping states to actions, parameterized by weights $\theta$
  • Training: Optimize $\theta$ to maximize expected cumulative reward $E[\sum_t \gamma^t r_t]$ using policy gradient methods

Advantages over classical control:

  • Learns nonlinear plasma response (no linearization assumptions)
  • Naturally multi-objective (all goals encoded in single reward)
  • Adapts to changing conditions (if trained with sufficient diversity)
  • Can discover non-intuitive control strategies

Disadvantages:

  • Requires extensive training data or high-fidelity simulator
  • Difficult to guarantee safety constraints (critical for a nuclear facility)
  • "Black box" behavior complicates regulatory approval
  • Sim-to-real transfer gap (simulator inaccuracies propagate to policy)

§7. The DeepMind-TCV Breakthrough and Its Limits

In February 2022, DeepMind (Alphabet) published a landmark paper demonstrating reinforcement learning control of a real tokamak plasma on the TCV (Tokamak à Configuration Variable) at EPFL, Switzerland.

What was achieved:

  • A single RL agent controlled 19 poloidal field coils simultaneously to maintain plasma shape, position, and current
  • The agent was trained entirely in simulation (LIUQE free-boundary equilibrium solver) and deployed on the real TCV without fine-tuning
  • Multiple plasma configurations were demonstrated:
    • Standard elongated plasmas
    • Droplet-shaped configurations
    • Simultaneous control of two separate plasmas (a first for RL control)
  • Control performance matched or exceeded the existing PID-based control system

Why this matters:

  1. Proof of concept for sim-to-real transfer: The RL agent worked on the real machine after training only in simulation. This validates the simulation-first approach for fusion control.
  2. Multi-objective capability: The agent maintained plasma shape, position, and current simultaneously — a task that previously required multiple independent controllers.
  3. Novel configurations: The agent discovered control strategies for configurations that had never been attempted with classical control (particularly the dual-plasma configuration).

What this does NOT mean:

  1. TCV is a small, low-power tokamak. Major radius 0.88 m, plasma current ~200 kA, pulse duration ~2 s. Scaling to ITER (6.2 m, 15 MA, 400+ s) introduces qualitatively new challenges: slower time scales, burning plasma dynamics, alpha particle heating feedback, and disruption consequences.

  2. TCV has no fusion neutrons. The control challenge of a D-T plasma — with alpha heating feedback, tritium burn control, and neutron-degraded diagnostics — is fundamentally different from the TCV demonstration.

  3. Safety constraints were not binding. On TCV, a failed control attempt costs a few seconds of experimental time. On ITER, a failed control event during D-T operation could cause machine damage costing months of repair and millions of dollars. The RL agent's training did not include hard safety constraints of this severity.

  4. The sim-to-real gap will be larger on ITER. TCV plasmas are well-modeled by existing equilibrium codes. ITER burning plasmas will exhibit phenomena (alpha particle-driven instabilities, ELM triggering dynamics, wall erosion evolution) that are not captured in current simulators.

The honest assessment of DeepMind-TCV:

It is a genuine breakthrough in applying AI to fusion control. It demonstrates that RL can learn effective plasma control policies from simulation and deploy them on real hardware. It opens a credible path toward AI-augmented control systems for next-generation devices.

It is not a solution to ITER or DEMO plasma control. The gap between TCV and a burning plasma reactor is comparable to the gap between a laboratory drone and a commercial airliner — the principles are related, but the engineering challenges are qualitatively different.

Part III: Accelerating Materials and Design

§8. AI for Material Discovery — The 20 dpa Shortcut?

Volume 4 established that the irradiation data gap is the single most consequential unknown in fusion engineering: structural materials have been tested to ~20 dpa with fusion-relevant He/dpa ratios, but power plant operation requires 50–100+ dpa.

Can AI help close this gap?

What AI CAN do for materials:

1. Accelerate candidate screening:

The design space for fusion structural alloys is vast. Consider a reduced-activation steel with potential alloying elements (Cr, W, V, Ta, Mn, Si, Ti, plus trace elements). Each element can vary in concentration, and the processing conditions (heat treatment, cold working) add more dimensions. Exhaustive experimental exploration of this space would take decades.

ML surrogate models — trained on existing mechanical property databases (tensile strength, hardness, DBTT, creep rate) — can predict properties for untested compositions and guide experimental selection. Approaches include:

  • Gaussian Process Regression: Provides uncertainty estimates alongside predictions, enabling Bayesian optimization of composition space
  • Random Forests / Gradient Boosted Trees: Robust, interpretable models for tabular alloy data
  • Graph Neural Networks: Can learn directly from crystal structure and composition, capturing physics of atomic interactions
  • Active learning loops: Model predicts → experiments validate most uncertain predictions → model updates → repeat

This approach can reduce the number of alloys that need experimental testing by 10–100×, accelerating candidate identification from decades to years.

2. Predict microstructural evolution:

Phase-field simulations and kinetic Monte Carlo models of radiation damage are computationally expensive (days to weeks per condition). ML surrogate models trained on these simulations can predict:

  • Void nucleation and growth rates
  • Precipitate evolution under irradiation
  • Segregation profiles at grain boundaries
  • Helium bubble size distributions

These surrogates enable rapid scanning of temperature, dose, and composition parameters.

3. Extract maximum information from limited data:

When irradiation experiments are expensive and slow (they are), statistical methods that squeeze information from small datasets become critical:

  • Transfer learning from fission irradiation data to fusion conditions
  • Physically informed priors that constrain predictions to obey known material science laws
  • Uncertainty quantification that honestly reports what the model doesn't know

What AI CANNOT do for materials:

The fundamental limitation: You cannot predict what you have not measured if the governing physics is unknown.

The behavior of RAFM steel above 20 dpa with 10–15 appm He/dpa is not merely "uncalculated" — the underlying damage mechanisms at this regime are not fully characterized. Helium-stabilized void nucleation at high concentrations, synergistic effects of He and H transmutation, and grain boundary embrittlement under high He loading may involve threshold effects or phase transitions that are not present in the lower-dose data.

An ML model trained on 0–20 dpa data will interpolate within that range effectively. It will extrapolate to 50–100 dpa poorly, because extrapolation into a regime with potentially different physics is not a machine learning problem — it is a physics problem that requires new experiments.

The honest timeline impact:

Materials task Without AI With AI Compression
Candidate alloy screening 10–15 years 3–5 years 60–70%
Microstructural modeling 5–10 years 2–4 years 50–60%
Irradiation experiment design 3–5 years 1–2 years 50–60%
Irradiation testing (0→50 dpa) 15–20 years 12–18 years 10–20%
Post-irradiation examination 5–8 years 3–5 years 30–40%

The rate-limiting step — irradiation testing — is barely compressible. Neutrons take time. DONES achieves 20–30 dpa/year in its high-flux zone. Getting to 50 dpa takes 2+ years of continuous irradiation regardless of how smart the experimental design is.

AI's value is in ensuring that the right materials are in the beam from day one, and that maximum information is extracted from every specimen.

§9. Surrogate Models for Neutronics and TBR

Volume 3 established that the TBR margin is near zero, with engineering losses systematically eroding the ideal TBR of 1.15–1.17 to an effective 1.03–1.05. Optimizing the blanket design to maximize effective TBR requires exploring a large parameter space:

  • Li-6 enrichment (30–90%)
  • Breeder pebble diameter (0.2–1.0 mm)
  • Beryllium multiplier fraction (30–60 vol%)
  • Structural fraction (10–25%)
  • Coolant channel geometry
  • First wall thickness
  • Blanket module dimensions
  • Port and penetration layout

Each parameter combination requires a full MCNP neutronics calculation — a Monte Carlo simulation that typically runs for 4–24 hours on a computing cluster to achieve adequate statistical precision.

The surrogate model approach:

  1. Generate training data: Run 1,000–10,000 MCNP calculations spanning the parameter space (using Latin Hypercube or Sobol sampling for efficient coverage).

  2. Train surrogate: Fit a neural network, Gaussian Process, or other ML model to map design parameters → TBR, heat deposition, tritium production rate, material damage.

  3. Optimize: Use the surrogate for rapid (millisecond) evaluation of millions of design candidates. Gradient-based or evolutionary optimization finds the Pareto front of maximum TBR vs. minimum structural fraction vs. acceptable thermal-hydraulic performance.

  4. Validate: Run full MCNP calculations on the predicted optimal designs to verify surrogate accuracy.

Demonstrated performance:

Study Surrogate type Training data Prediction error (TBR) Speedup
Bonate et al. (2023) Neural network 5,000 MCNP runs ±0.005 (0.4%) ~10⁶×
Fischer et al. (2022) Gaussian Process 2,000 MCNP runs ±0.008 (0.7%) ~10⁵×
Palermo et al. (2021) Random Forest 3,000 MCNP runs ±0.01 (0.9%) ~10⁵×

A prediction error of ±0.005 in TBR is remarkable — comparable to the Monte Carlo statistical uncertainty of the MCNP calculations themselves. This means the surrogate is essentially as accurate as the physics code it was trained on.

What this enables:

  • Full-blanket optimization in hours instead of years
  • Sensitivity analysis across all design parameters simultaneously (Vol.3, §16 Monte Carlo analysis could use surrogates for even richer modeling)
  • Real-time design exploration during engineering reviews
  • Robustness analysis: How does TBR degrade under manufacturing tolerances?

Limitation: The surrogate is only as good as the underlying physics code. If MCNP's nuclear data libraries have systematic errors for certain reactions at 14 MeV, the surrogate faithfully reproduces those errors. AI amplifies the power of the physics code but does not improve its fundamental accuracy.

§10. Stellarator Optimization — AI's Greatest Fusion Success

If there is a single domain where AI has unambiguously transformed fusion research, it is stellarator design optimization.

The stellarator design problem:

A stellarator creates its confining magnetic field entirely from external coils (unlike a tokamak, which relies on a plasma current). The absence of plasma current eliminates disruptions — a massive safety advantage — but creates an immensely difficult coil design problem.

The magnetic field must satisfy simultaneously:

  • Quasi-symmetry or quasi-isodynamicity for good particle confinement
  • Low neoclassical transport (particles should not drift out of the device)
  • MHD stability at the operating beta
  • Acceptable coil complexity (the coils must be physically constructable)
  • Adequate plasma-coil separation (for blanket and maintenance access)
  • Acceptable magnetic field errors from finite coil dimensions

The design space is the shape of the magnetic field in 3D — effectively an infinite-dimensional optimization problem. Wendelstein 7-X, the world's largest stellarator, was designed using 1990s optimization tools that explored a tiny fraction of the feasible design space.

AI-driven optimization:

Modern stellarator optimization uses:

  • Adjoint methods for gradient computation (efficient sensitivity analysis)
  • Bayesian optimization for global search with expensive function evaluations
  • Evolutionary algorithms (genetic algorithms, particle swarm) for multi-objective Pareto fronts
  • Neural network surrogates for rapid evaluation of equilibrium, stability, and transport
  • Automatic differentiation through the physics codes (enabled by JAX, PyTorch backends)

The SIMSOPT framework (developed at PPPL) and related tools have enabled:

  • Discovery of stellarator configurations with neoclassical transport 100× better than W7-X
  • Identification of "precise quasi-symmetry" configurations that were not known to exist
  • Optimization of coil shapes for manufacturability and maintenance access simultaneously
  • Design of configurations that minimize turbulent transport (a much harder objective)

Concrete impact:

The QUASAR configuration — a quasi-axisymmetric stellarator optimized with modern tools — achieves confinement quality comparable to tokamaks without plasma current. Previous generations of stellarator design could not achieve this because the optimization tools could not navigate the design space effectively.

Why this is AI's greatest fusion success:

  1. The problem is well-posed: The physics codes (VMEC, SPEC, BOOZ_XFORM) are mature and accurate. The design space is high-dimensional but well-defined. This is the ideal setting for optimization.

  2. The impact is measurable: Confinement quality improvement of 10–100× over previous designs, achieved in years rather than decades of manual iteration.

  3. The result is actionable: Optimized stellarator configurations can be built with current engineering capabilities. Several next-generation stellarator proposals (QUASAR, CFQS, Muse) are direct products of AI-assisted optimization.

  4. The limitation is clear: AI optimized within the design space defined by existing physics. It did not discover new physics. It found better solutions to a well-understood problem.

§11. Reactor Design Space Exploration

Beyond individual subsystem optimization, AI enables system-level fusion reactor design exploration — simultaneously varying plasma parameters, magnet design, blanket configuration, and power conversion to find globally optimal designs.

The system code approach:

Fusion system codes (PROCESS, ARIES-AT, HELIAS) model an entire reactor as a set of coupled algebraic and differential equations: plasma physics (0D or 1D), magnet engineering, neutronics, thermal-hydraulics, power balance, cost estimation, and maintenance scheduling. A single run takes seconds to minutes.

AI-enhanced system codes:

  • Use surrogate models for the most expensive sub-calculations (neutronics, structural analysis)
  • Employ multi-objective optimization to find Pareto fronts (e.g., minimum cost vs. maximum availability vs. minimum tritium consumption)
  • Perform uncertainty propagation through the entire design chain

Example: DEMO design space (EU-DEMO parameters from Vol.4):

The design space includes:

  • Major radius: 7–11 m
  • Aspect ratio: 2.5–4.0
  • Magnetic field on axis: 4–7 T
  • Plasma current: 12–20 MA
  • Fusion power: 1,500–3,000 MW
  • TBR: 1.03–1.10
  • Blanket type: HCPB vs. WCLL
  • Divertor concept: W monoblock vs. liquid metal
  • Maintenance scheme: port-based vs. vertical

Each combination implies different tritium consumption, neutron wall loading, material dose accumulation, thermal efficiency, capital cost, and electricity cost.

Exhaustive exploration of this space with full-fidelity codes would require millions of evaluations — impractical even on modern supercomputers. AI surrogate-assisted optimization reduces this to thousands of full evaluations plus millions of surrogate evaluations, enabling identification of robust design points in weeks rather than years.

The PROCESS code + AI workflow (under development at CCFE/UKAEA):

  1. Run 5,000 PROCESS evaluations spanning design space
  2. Train multi-output neural network surrogate
  3. Perform Bayesian optimization on surrogate to find optimal designs
  4. Validate top 100 candidates with full PROCESS runs
  5. Select robust designs that perform well across uncertainty ranges

This workflow has identified design points with 15–20% lower capital cost than previous reference designs while maintaining all engineering constraints — a significant result for a technology whose economic viability is uncertain.

Part IV: The Honest Assessment

§12. Disruption Prediction Demonstration (Python)

The following code demonstrates a simplified disruption prediction pipeline using synthetic data that mimics the statistical structure of real tokamak diagnostics. This is for illustrative purposes — real disruption predictors use proprietary machine-specific diagnostic data.

"""
Disruption Prediction Demonstration — Simplified ML Pipeline
Nuclear Fusion Vol.5, §12
Author: Dosanko Tousan | AI Partner: Claude (Anthropic)
License: MIT

NOTE: Uses synthetic data mimicking tokamak diagnostic statistics.
Real disruption predictors use machine-specific data from JET, DIII-D, etc.
"""

import numpy as np
import matplotlib.pyplot as plt
from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import (roc_curve, auc, precision_recall_curve,
                              confusion_matrix, classification_report)
from sklearn.preprocessing import StandardScaler

np.random.seed(42)

# ============================================================
# Synthetic tokamak diagnostic data generation
# ============================================================
def generate_tokamak_data(n_discharges=2000, disruption_fraction=0.15):
    """
    Generate synthetic diagnostic time-slice features.
    Features mimic real tokamak signals that precede disruptions.
    """
    n_disruptions = int(n_discharges * disruption_fraction)
    n_safe = n_discharges - n_disruptions

    features_safe = np.column_stack([
        np.random.normal(0.7, 0.15, n_safe),    # beta_N / beta_limit (safe: ~0.7)
        np.random.normal(0.6, 0.12, n_safe),     # n_e / n_Greenwald (safe: ~0.6)
        np.random.exponential(0.1, n_safe),       # locked mode amplitude (safe: low)
        np.random.normal(0.3, 0.08, n_safe),      # radiated power fraction (safe: ~0.3)
        np.random.normal(1.2, 0.15, n_safe),      # internal inductance li (safe: ~1.2)
        np.random.normal(0.05, 0.02, n_safe),     # dI_p/dt / I_p (safe: ~0)
        np.random.normal(2.0, 0.3, n_safe),       # q95 (safe: ~2.0 above limit)
        np.random.normal(0.0, 0.5, n_safe),       # vertical position z [cm] (safe: ~0)
    ])

    features_disrupt = np.column_stack([
        np.random.normal(0.92, 0.08, n_disruptions),  # beta_N near limit
        np.random.normal(0.85, 0.10, n_disruptions),  # density near Greenwald
        np.random.exponential(0.5, n_disruptions) + 0.3,  # locked mode growing
        np.random.normal(0.6, 0.12, n_disruptions),   # high radiated fraction
        np.random.normal(1.6, 0.20, n_disruptions),   # elevated li
        np.random.normal(-0.15, 0.08, n_disruptions), # current dropping
        np.random.normal(1.5, 0.2, n_disruptions),    # q95 dropping toward limit
        np.random.normal(2.0, 1.5, n_disruptions),    # vertical displacement
    ])

    X = np.vstack([features_safe, features_disrupt])
    y = np.concatenate([np.zeros(n_safe), np.ones(n_disruptions)])

    shuffle_idx = np.random.permutation(len(y))
    return X[shuffle_idx], y[shuffle_idx]

feature_names = ['β_N/β_limit', 'n_e/n_GW', 'Locked Mode',
                 'P_rad/P_tot', 'l_i', 'dI_p/dt/I_p', 'q_95', 'z_pos [cm]']

X, y = generate_tokamak_data(n_discharges=5000, disruption_fraction=0.15)

# ============================================================
# Train/test split and scaling
# ============================================================
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.3, random_state=42, stratify=y)

scaler = StandardScaler()
X_train_s = scaler.fit_transform(X_train)
X_test_s = scaler.transform(X_test)

# ============================================================
# Models
# ============================================================
models = {
    'Random Forest': RandomForestClassifier(
        n_estimators=200, max_depth=12, min_samples_leaf=5, random_state=42),
    'Gradient Boosted Trees': GradientBoostingClassifier(
        n_estimators=200, max_depth=6, learning_rate=0.1, random_state=42),
}

results = {}
for name, model in models.items():
    model.fit(X_train_s, y_train)
    y_prob = model.predict_proba(X_test_s)[:, 1]
    y_pred = model.predict(X_test_s)
    fpr, tpr, _ = roc_curve(y_test, y_prob)
    roc_auc = auc(fpr, tpr)
    precision, recall, _ = precision_recall_curve(y_test, y_prob)
    results[name] = {
        'fpr': fpr, 'tpr': tpr, 'roc_auc': roc_auc,
        'precision': precision, 'recall': recall,
        'y_pred': y_pred, 'y_prob': y_prob, 'model': model
    }
    print(f"\n{name}:")
    print(f"  ROC AUC: {roc_auc:.4f}")
    print(classification_report(y_test, y_pred, target_names=['Safe', 'Disruption']))

# ============================================================
# Figures
# ============================================================
fig, axes = plt.subplots(1, 3, figsize=(18, 5.5), dpi=150)
colors = {'Random Forest': '#3498db', 'Gradient Boosted Trees': '#e74c3c'}

# Panel 1: ROC curves
ax1 = axes[0]
for name, res in results.items():
    ax1.plot(res['fpr'], res['tpr'], color=colors[name], linewidth=2.5,
             label=f'{name} (AUC={res["roc_auc"]:.3f})')
ax1.plot([0, 1], [0, 1], 'k--', linewidth=1, alpha=0.5)
ax1.set_xlabel('False Positive Rate', fontsize=12)
ax1.set_ylabel('True Positive Rate', fontsize=12)
ax1.set_title('ROC Curves — Disruption Prediction', fontsize=13, fontweight='bold')
ax1.legend(fontsize=10)
ax1.grid(True, alpha=0.3)

# Panel 2: Feature importance (best model)
ax2 = axes[1]
best_model = results['Gradient Boosted Trees']['model']
importances = best_model.feature_importances_
sorted_idx = np.argsort(importances)
ax2.barh(range(len(importances)), importances[sorted_idx],
         color='#2ecc71', edgecolor='white', height=0.6)
ax2.set_yticks(range(len(feature_names)))
ax2.set_yticklabels([feature_names[i] for i in sorted_idx], fontsize=10)
ax2.set_xlabel('Feature Importance', fontsize=12)
ax2.set_title('What Drives Disruption Risk?', fontsize=13, fontweight='bold')
ax2.grid(True, alpha=0.3, axis='x')

# Panel 3: Precision-Recall
ax3 = axes[2]
for name, res in results.items():
    ax3.plot(res['recall'], res['precision'], color=colors[name], linewidth=2.5,
             label=name)
ax3.axhline(y=0.95, color='gray', linestyle='--', linewidth=1.5, alpha=0.7)
ax3.text(0.05, 0.96, 'Target: 95% precision', fontsize=9, color='gray')
ax3.set_xlabel('Recall (True Positive Rate)', fontsize=12)
ax3.set_ylabel('Precision', fontsize=12)
ax3.set_title('Precision-Recall Tradeoff', fontsize=13, fontweight='bold')
ax3.legend(fontsize=10)
ax3.set_xlim(0, 1.05)
ax3.set_ylim(0.5, 1.02)
ax3.grid(True, alpha=0.3)

plt.suptitle('§12: ML Disruption Prediction (Synthetic Tokamak Data)',
             fontsize=14, fontweight='bold', y=1.02)
plt.tight_layout()
plt.savefig('fig1_disruption_prediction.png', bbox_inches='tight', facecolor='white')
plt.close()

print("\nFigure 1 saved: fig1_disruption_prediction.png")
Enter fullscreen mode Exit fullscreen mode

§13. Material Screening Surrogate (Python) 

"""
Material Screening Surrogate — RAFM Steel Property Prediction
Nuclear Fusion Vol.5, §13
Author: Dosanko Tousan | AI Partner: Claude (Anthropic)
License: MIT

Demonstrates the surrogate model concept for rapid material screening.
Uses synthetic alloy composition → property mapping.
"""

import numpy as np
import matplotlib.pyplot as plt
from sklearn.ensemble import GradientBoostingRegressor
from sklearn.model_selection import train_test_split
from sklearn.metrics import r2_score, mean_absolute_error

np.random.seed(42)

# ============================================================
# Synthetic RAFM alloy database
# ============================================================
def generate_alloy_database(n_alloys=3000):
    """
    Generate synthetic RAFM alloy compositions and properties.
    Composition ranges based on real RAFM design space.
    Property models are simplified physics-inspired functions.
    """
    Cr = np.random.uniform(7.0, 12.0, n_alloys)    # wt%
    W  = np.random.uniform(0.5, 3.0, n_alloys)     # wt%
    V  = np.random.uniform(0.05, 0.40, n_alloys)   # wt%
    Ta = np.random.uniform(0.02, 0.25, n_alloys)   # wt%
    Mn = np.random.uniform(0.1, 0.8, n_alloys)     # wt%
    C  = np.random.uniform(0.06, 0.18, n_alloys)   # wt%
    N  = np.random.uniform(0.01, 0.06, n_alloys)   # wt%

    X = np.column_stack([Cr, W, V, Ta, Mn, C, N])

    # Simplified property models (inspired by metallurgical principles)
    yield_strength = (350 + 25*Cr + 80*W + 200*V + 150*Ta + 1500*C
                      + 30*Mn + 800*N + np.random.normal(0, 20, n_alloys))

    dbtt_unirrad = (-120 + 5*Cr - 10*W + 50*V - 30*Ta + 200*C
                    + 10*Mn - 100*N + np.random.normal(0, 10, n_alloys))

    # DBTT shift after 10 dpa at 300°C (depends on composition)
    dbtt_shift_10dpa = (100 + 8*Cr + 15*W - 40*V - 20*Ta + 300*C
                        + 5*Mn + np.random.normal(0, 15, n_alloys))

    # Creep rupture time at 550°C, 150 MPa [hours]
    log_creep_life = (4.0 + 0.05*Cr + 0.3*W + 0.5*V + 0.4*Ta + 2.0*C
                      - 0.1*Mn + np.random.normal(0, 0.2, n_alloys))
    creep_life = 10**log_creep_life

    return X, yield_strength, dbtt_unirrad, dbtt_shift_10dpa, creep_life

comp_names = ['Cr', 'W', 'V', 'Ta', 'Mn', 'C', 'N']
prop_names = ['Yield Strength [MPa]', 'DBTT (unirrad) [°C]',
              'ΔDBTT (10dpa, 300°C) [°C]', 'Creep Life [hours]']

X, ys, dbtt, dbtt_shift, creep = generate_alloy_database(5000)

# ============================================================
# Train surrogates for each property
# ============================================================
targets = [ys, dbtt, dbtt_shift, np.log10(creep)]
target_labels = prop_names.copy()
target_labels[3] = 'log₁₀(Creep Life [hours])'

fig, axes = plt.subplots(2, 2, figsize=(14, 11), dpi=150)

for idx, (y_data, label, ax) in enumerate(zip(targets, target_labels, axes.ravel())):
    X_tr, X_te, y_tr, y_te = train_test_split(X, y_data, test_size=0.2, random_state=42)

    model = GradientBoostingRegressor(
        n_estimators=200, max_depth=5, learning_rate=0.1, random_state=42)
    model.fit(X_tr, y_tr)
    y_pred = model.predict(X_te)

    r2 = r2_score(y_te, y_pred)
    mae = mean_absolute_error(y_te, y_pred)

    ax.scatter(y_te, y_pred, alpha=0.3, s=8, color='#3498db')
    lims = [min(y_te.min(), y_pred.min()), max(y_te.max(), y_pred.max())]
    ax.plot(lims, lims, 'r--', linewidth=2, label='Perfect prediction')
    ax.set_xlabel(f'True {label}', fontsize=11)
    ax.set_ylabel(f'Predicted {label}', fontsize=11)
    ax.set_title(f'{prop_names[idx]}\nR² = {r2:.4f}, MAE = {mae:.2f}',
                 fontsize=12, fontweight='bold')
    ax.legend(fontsize=9)
    ax.grid(True, alpha=0.3)

    print(f"{prop_names[idx]}: R² = {r2:.4f}, MAE = {mae:.2f}")

plt.suptitle('§13: Material Screening Surrogate — RAFM Steel Properties',
             fontsize=14, fontweight='bold', y=1.01)
plt.tight_layout()
plt.savefig('fig2_material_surrogate.png', bbox_inches='tight', facecolor='white')
plt.close()
print("\nFigure 2 saved: fig2_material_surrogate.png")

# ============================================================
# Demonstrate: find optimal composition for fusion
# ============================================================
print("\n" + "="*60)
print("Optimal RAFM Composition Search (surrogate-guided)")
print("="*60)
print("Objective: Maximize yield strength")
print("         + Minimize DBTT shift after irradiation")
print("         + Maximize creep life")
print("Constraints: Yield > 500 MPa, DBTT_shift < 150°C\n")

# Train all models on full dataset for screening
models_full = []
for y_data in targets:
    m = GradientBoostingRegressor(n_estimators=200, max_depth=5,
                                   learning_rate=0.1, random_state=42)
    m.fit(X, y_data)
    models_full.append(m)

# Screen 100,000 random compositions
n_screen = 100000
X_screen = np.column_stack([
    np.random.uniform(7.0, 12.0, n_screen),
    np.random.uniform(0.5, 3.0, n_screen),
    np.random.uniform(0.05, 0.40, n_screen),
    np.random.uniform(0.02, 0.25, n_screen),
    np.random.uniform(0.1, 0.8, n_screen),
    np.random.uniform(0.06, 0.18, n_screen),
    np.random.uniform(0.01, 0.06, n_screen),
])

pred_ys = models_full[0].predict(X_screen)
pred_dbtt_shift = models_full[2].predict(X_screen)
pred_creep = 10**models_full[3].predict(X_screen)

# Filter and rank
mask = (pred_ys > 500) & (pred_dbtt_shift < 150)
filtered_idx = np.where(mask)[0]

if len(filtered_idx) > 0:
    # Combined score (normalized)
    score = (pred_ys[filtered_idx] / pred_ys[filtered_idx].max()
             - pred_dbtt_shift[filtered_idx] / pred_dbtt_shift[filtered_idx].max()
             + np.log10(pred_creep[filtered_idx]) / np.log10(pred_creep[filtered_idx]).max())
    best_idx = filtered_idx[np.argmax(score)]

    print(f"Screened {n_screen:,} compositions, {len(filtered_idx):,} passed constraints")
    print(f"\nBest candidate composition:")
    for name, val in zip(comp_names, X_screen[best_idx]):
        print(f"  {name}: {val:.3f} wt%")
    print(f"\nPredicted properties:")
    print(f"  Yield Strength: {pred_ys[best_idx]:.0f} MPa")
    print(f"  DBTT shift (10 dpa): {pred_dbtt_shift[best_idx]:.0f} °C")
    print(f"  Creep life (550°C): {pred_creep[best_idx]:.0f} hours")
    print(f"\n→ This candidate should be prioritized for experimental validation.")
Enter fullscreen mode Exit fullscreen mode

§14. What AI Cannot Do — The Honest Section

This is the most important section in Volume 5. The fusion-AI field is subject to substantial hype, and distinguishing between demonstrated capability and aspirational projection is essential for sound investment decisions.

AI cannot replace experiments that have not been performed.

This is the single most important sentence in this volume. Machine learning is, at its core, a technology for interpolation and pattern recognition within the distribution of its training data. It excels at:

  • Finding patterns in existing data that humans miss
  • Interpolating between measured conditions
  • Optimizing within well-characterized design spaces
  • Accelerating simulations through surrogate modeling

It struggles fundamentally with:

  • Extrapolating beyond the training data distribution
  • Predicting phenomena governed by physics that is not represented in training data
  • Discovering qualitatively new failure modes
  • Replacing physical validation of safety-critical systems

Specific fusion tasks where AI hype exceeds demonstrated capability:

1. "AI will predict material behavior at 50–100 dpa"

No. AI can predict material behavior within the regime of existing data (~0–20 dpa for fusion-relevant He/dpa). Beyond 20 dpa, the governing damage mechanisms may change qualitatively (helium bubble coalescence, grain boundary decohesion, phase instability). These phenomena are not in the training data because they have not been observed under fusion conditions. An ML model that predicts "smooth extrapolation" from 20 dpa behavior would be precisely wrong if a threshold effect occurs at 30 dpa.

2. "AI will solve the disruption problem on ITER"

Partially. AI can significantly improve disruption prediction accuracy and warning time, based on transfer learning from existing tokamaks. But ITER's disruption dynamics involve phenomena (alpha-heated plasmas, high-Z metallic walls, high-current runaway electron beams) that do not exist in current training data. The first few hundred ITER discharges will be the training data — they cannot be avoided by pre-existing AI.

3. "AI will design the optimal fusion reactor"

AI can optimize within the design space defined by current physics understanding. It cannot design around physics that is not yet understood. The breeding blanket TBR, for example, depends on nuclear cross sections at 14 MeV that carry 5–10% uncertainty for some reactions. AI optimization finds the best design given current cross section data — but if the data is wrong, the design is wrong, and AI cannot detect the discrepancy.

4. "AI will accelerate fusion by a decade"

The overall program timeline is limited by the slowest serial dependency: irradiation testing, which requires physical neutrons at physical rates. DONES achieves ~25 dpa/year. Getting 50 dpa of validated material data takes ~2 years of DONES operation regardless of AI. AI can ensure the right samples are in the beam and maximum information is extracted, but it cannot make neutrons travel faster.

What AI CAN legitimately accelerate (summary):

Task Mechanism Realistic compression
Disruption prediction development Transfer learning, data fusion 40–50%
Stellarator design Surrogate-assisted optimization 50–75%
Material candidate screening Surrogate models, active learning 60–70%
Blanket TBR optimization Neural network surrogates for MCNP 50–60%
Reactor system design Multi-objective surrogate optimization 60–70%
Irradiation testing Optimal experimental design 10–20%
Engineering integration Limited (physical prototype required) 5–10%

The aggregate impact is a program timeline compression of approximately 25%, or 5–7 years on a 25–30 year horizon. This is meaningful — it could be the difference between making and missing the tritium supply window — but it is not the order-of-magnitude acceleration that some advocates claim.

The danger of AI hype in fusion:

If policymakers believe that AI can replace physical validation, they may underinvest in the critical experimental infrastructure (DONES, ITER TBMs, material irradiation programs) that no amount of computation can substitute. The most dangerous failure mode for fusion AI is not that it doesn't work — it's that it works well enough in the regimes where data exists to create false confidence about regimes where data doesn't exist.

§15. Timeline Compression Assessment

Integrating the analysis from all preceding sections, we can now estimate the overall impact of AI on the fusion program timeline.

The critical path (without AI):

2025 ─── Material candidates selected ───── 2030
          ↓
2030 ─── DONES first beam ──────────────── 2032
          ↓
2032 ─── DONES 20 dpa data ────────────── 2035
          ↓
2035 ─── ITER first plasma ────────────── 2035
          ↓
2035 ─── DONES 50 dpa data ────────────── 2040
          ↓
2039 ─── ITER D-T operations begin ─────── 2042
          ↓
2042 ─── TBM data + material qualification ── 2048
          ↓
2048 ─── DEMO design freeze ───────────── 2048
          ↓
2050 ─── DEMO first plasma ────────────── 2055
          ↓
2060 ─── First fusion electricity ──────── 2065
Enter fullscreen mode Exit fullscreen mode

The critical path (with AI, optimistic but realistic):

2025 ─── AI-screened material candidates ── 2027 (saved ~3 years)
          ↓
2030 ─── DONES first beam ──────────────── 2032 (unchanged: hardware)
          ↓
2032 ─── DONES 20 dpa + AI analysis ────── 2034 (saved ~1 year)
          ↓
2035 ─── ITER first plasma + AI control ── 2035 (unchanged: hardware)
          ↓
2035 ─── DONES 50 dpa + AI extrapolation ─ 2038 (saved ~2 years)
          ↓
2039 ─── ITER D-T + AI disruption predict ── 2040 (saved ~2 years)
          ↓
2040 ─── TBM data + AI-optimized blanket ── 2044 (saved ~4 years)
          ↓
2044 ─── DEMO design freeze ───────────── 2044 (saved ~4 years)
          ↓
2047 ─── DEMO first plasma ────────────── 2050 (saved ~5 years)
          ↓
2055 ─── First fusion electricity ──────── 2060 (saved ~5 years)
Enter fullscreen mode Exit fullscreen mode

Net compression: approximately 5 years on the overall timeline.

This 5-year acceleration is not uniformly distributed. The largest gains come from:

  1. Material screening (AI selects better candidates sooner)
  2. Blanket optimization (AI-assisted TBR design finds optimal configuration faster)
  3. Disruption avoidance (AI enables higher-performance operation sooner)

The smallest gains come from:

  1. DONES/ITER hardware construction (AI cannot build accelerators or tokamaks faster)
  2. Material irradiation (neutrons don't care about algorithms)
  3. Engineering integration (physical prototypes are irreplaceable)

The 5-year question:

Is 5 years significant? Consider:

  • The tritium cliff (Vol.3): inventory crosses the DEMO threshold around 2045–2050
  • Without AI: DEMO design freeze in 2048 → may be too late
  • With AI: DEMO design freeze in 2044 → feasible margin

Five years may be the difference between a fusion program that catches the tritium supply window and one that misses it. In this context, AI's contribution is not merely incremental — it may be existentially necessary.

§16. Decision Matrix

AI-Fusion Investment Area TRL Demonstrated Impact Risk Priority
Disruption prediction (cross-machine) 5–6 High (>95% TPR on home machine) Transfer to ITER unvalidated ★★★★☆
RL plasma control 4–5 Moderate (TCV demo) Gap to burning plasma large ★★★☆☆
Stellarator optimization 6–7 Very high (10–100× improvement) Well-posed problem, low risk ★★★★★
Neutronics surrogate (TBR) 5–6 High (±0.5% accuracy) Limited by underlying code ★★★★☆
Material screening 4–5 Moderate (limited by data) Extrapolation risk high ★★★★☆
Material property prediction (high dose) 2–3 Low (no fusion data) False confidence risk ★★☆☆☆
Reactor system optimization 4–5 Moderate (15–20% cost reduction) Design space well-characterized ★★★★☆
Real-time burn control 2 None (no burning plasma) Entirely forward-looking ★★☆☆☆
Foundation models for fusion 3 Low (early research) Promising but unproven ★★★☆☆

Reading the matrix:

  • ★★★★★: Invest now, demonstrated impact, well-posed problem
  • ★★★★☆: Invest now, high potential, some risks to manage
  • ★★★☆☆: Invest selectively, promising but unproven at scale
  • ★★☆☆☆: Research-stage, beware hype, manage expectations
  • ★☆☆☆☆: Not recommended at current maturity

The single highest-return investment in fusion AI is stellarator optimization. The problem is well-posed, the impact is demonstrated, and the results are directly actionable. If fusion's future includes stellarators (disruption-free by design), AI has already contributed more to that path than to any other.

§17. Conclusions and Forward Look

This volume has assessed the role of AI in fusion research with a commitment to honesty over hype.

What AI has already achieved:

  • Disruption prediction at >95% accuracy on multiple machines
  • Stellarator optimization yielding 10–100× better confinement configurations
  • Real-time RL plasma control demonstrated on TCV
  • Neutronics surrogate models replacing hours of MCNP with milliseconds of inference

What AI will likely achieve in the next decade:

  • Cross-machine disruption prediction validated on JT-60SA and early ITER plasmas
  • Material screening that identifies superior RAFM compositions for DONES irradiation
  • Blanket TBR optimization that pushes effective TBR from 1.03 toward 1.07
  • System-level reactor design optimization that reduces DEMO capital cost by 15–20%

What AI will not achieve:

  • Replacement of material irradiation experiments
  • Prediction of unknown failure modes at >20 dpa
  • Elimination of the tritium cliff through computation
  • Guaranteed disruption avoidance on ITER from day one

The central message:

AI is a force multiplier, not a magic wand. It amplifies the value of every experiment, every simulation, and every design iteration. In a program where the binding constraint is time — the tritium cliff, the DONES schedule, the DEMO construction window — a 25% timeline compression is the difference between success and failure.

The fusion program needs AI. It also needs DONES, ITER, and the engineering test facilities that no algorithm can replace. The investment strategy must be "both/and," not "either/or."

Connection to the series:

Volumes 1–4 established what fusion must achieve (physics) and what stands in the way (engineering). This volume established what AI can and cannot contribute to closing the gap. The remaining volumes address the question Volumes 3–4 raised implicitly: What if we choose a different path?

Vol.6 examines whether advanced fuels (D-³He, p-¹¹B) can eliminate the tritium and materials crises entirely — at the cost of far more demanding ignition conditions. Vol.8 asks whether fusion propulsion can bypass the power plant paradigm altogether. The answers require the quantitative rigor established in Volumes 1–5 and the AI tools assessed in this volume.

The path to Valkyrie runs through all of them.

References

  1. J. Kates-Harbeck, A. Svyatkovskiy, and W. Tang, "Predicting disruptive instabilities in controlled fusion plasmas through deep learning," Nature, vol. 568, pp. 526–531 (2019).

  2. C. Rea et al., "Disruption prediction investigations using machine learning tools on DIII-D and Alcator C-Mod," Plasma Physics and Controlled Fusion, vol. 60, 084004 (2018).

  3. J. Degrave et al., "Magnetic control of tokamak plasmas through deep reinforcement learning," Nature, vol. 602, pp. 414–419 (2022). [DeepMind-TCV paper]

  4. M. Landreman and E. Paul, "Magnetic fields with precise quasisymmetry for plasma confinement," Physical Review Letters, vol. 128, 035001 (2022).

  5. F. Wechsung et al., "Precise stellarator quasi-symmetry can be achieved with electromagnetic coils," Proceedings of the National Academy of Sciences, vol. 119, e2202084119 (2022).

  6. M. Shimada et al., "SIMSOPT: A flexible framework for stellarator optimization," Journal of Open Source Software, vol. 6, 3525 (2021).

  7. P. C. de Vries et al., "Multi-machine analysis of JET disruption prediction," Plasma Physics and Controlled Fusion, vol. 63, 054002 (2021).

  8. A. Ho et al., "Neural network surrogate of QuaLiKiz for multi-scale integrated modelling of tokamak plasmas," Physics of Plasmas, vol. 28, 032305 (2021).

  9. G. Dong et al., "Machine learning surrogate models for prediction of point defect vibrational free energies," Journal of Nuclear Materials, vol. 544, 152693 (2021).

  10. J. Citrin et al., "Real-time capable first-principle based modelling of tokamak turbulent transport," Nuclear Fusion, vol. 55, 092001 (2015).

  11. TAE Technologies, "Optometrist algorithm for plasma optimization," Nature Communications (2020).

  12. S. Creely et al., "Overview of the SPARC tokamak," Journal of Plasma Physics, vol. 86, 865860502 (2020).

  13. R. Kembleton et al., "AI-enhanced PROCESS system code for fusion reactor design," Fusion Engineering and Design (2023).

  14. D. Stork et al., "Materials R&D for a timely DEMO: key findings and recommendations of the EU Roadmap Materials Assessment Group," Fusion Engineering and Design, vol. 89, pp. 1586–1594 (2014).

This volume was written by Dosanko Tousan with Claude (Anthropic) as AI partner.
The honest section (§14) was written first. Everything else was written to deserve it.
AI wrote this article about AI in fusion. The irony is not lost on us.

fig5_material_surrogate.png

Top comments (0)