--- id: ai-for-nuclear-fusion title: "AI for Nuclear Fusion: Plasma Control, Disruption Prediction, and Accelerated Discovery" schema_type: article category: ai language: en confidence: high last_verified: "2026-05-24" created_date: "2026-05-24" generation_method: ai_assisted ai_models: - claude-4.5-sonnet derived_from_human_seed: true conflict_of_interest: none_declared is_live_document: false data_period: static completeness: 0.85 atomic_facts: - id: af-ai-for-nuclear-fusion-1 statement: >- Google DeepMind (Nature, February 2022) demonstrated the first application of deep reinforcement learning to control magnetic confinement of fusion plasma in the TCV tokamak. The controller manipulated 19 magnetic coils to autonomously produce and sustain advanced plasma configurations, including "droplet" and "snowflake" divertor shapes previously achievable only through extensive manual engineering. source_title: Degrave et al., Nature (2022) -- doi:10.1038/s41586-021-04301-9 source_url: https://www.nature.com/articles/s41586-021-04301-9 confidence: high - id: af-ai-for-nuclear-fusion-2 statement: >- AI-driven disruption prediction is critical for ITER and large tokamaks. Deep learning models (LSTM, CNN-LSTM, ensemble methods) trained on multi-diagnostic plasma signals can predict disruptive instabilities 30-100ms before occurrence. The DisruptionBench platform (Springer 2025) standardized evaluation across multiple tokamaks, with best models achieving >90% true positive rate at <5% false positive rate. source_title: DisruptionBench (Springer 2025) -- doi:10.1007/s10894-025-00495-2 / US DOE (2025) source_url: https://link.springer.com/article/10.1007/s10894-025-00495-2 confidence: high primary_sources: - id: ps-ai-for-nuclear-fusion-1 title: Magnetic control of tokamak plasmas through deep reinforcement learning type: academic_paper year: 2022 institution: Nature / Google DeepMind doi: 10.1038/s41586-021-04301-9 url: https://www.nature.com/articles/s41586-021-04301-9 - id: ps-ai-for-nuclear-fusion-2 title: "DisruptionBench and Complimentary New Models: Systematically Evaluate ML-Driven Disruption Prediction" type: academic_paper year: 2025 institution: Journal of Fusion Energy (Springer) doi: 10.1007/s10894-025-00495-2 url: https://link.springer.com/article/10.1007/s10894-025-00495-2 known_gaps: - Transfer learning between tokamaks -- models trained on one device rarely generalize to others - Real-time interpretable AI -- operators need to understand why a disruption is predicted before acting disputed_statements: [] secondary_sources: - title: Magnetic Control of Tokamak Plasmas Through Deep Reinforcement Learning (TCV) type: journal_article year: 2022 authors: - Degrave, Jonas - Frezal, Federico - Foguel, Sebastien - et al. institution: Google DeepMind / EPFL / Nature url: https://www.nature.com/articles/s41586-021-04301-9 - title: "AI for Fusion Energy: A Comprehensive Survey of ML Applications in Plasma Physics and Fusion Reactor Design" type: survey_paper year: 2024 authors: - multiple institution: Nature Reviews Physics url: https://doi.org/10.1038/s42254-024-00734-5 - title: Avoiding Tokamak Tearing Instabilities with Deep Reinforcement Learning (ITER) type: journal_article year: 2024 authors: - Seo, Jaemin - Kim, SangKyeun - Jalalvand, Azarakhsh - et al. institution: Princeton Plasma Physics Lab / Nature url: https://www.nature.com/articles/s41586-024-07024-9 - title: "IAEA Report: Artificial Intelligence for Accelerating Fusion Energy Development" type: report year: 2024 authors: - IAEA institution: International Atomic Energy Agency url: https://www.iaea.org/topics/fusion updated: "2026-05-24" --- ## TL;DR AI is accelerating the path to commercial fusion energy by solving two of the hardest problems in plasma physics: real-time control of magnetically confined 100-million-degree plasma and prediction of dangerous instabilities (disruptions) that can damage reactor walls. From DeepMind's deep RL plasma controller to disruption prediction systems deployed on ITER, machine learning is becoming an essential tool in the fusion engineer's toolkit. ## Core Explanation Nuclear fusion promises virtually unlimited clean energy if we can confine and control a plasma at temperatures exceeding 100 million degrees Celsius. The leading approach, the tokamak (a doughnut-shaped magnetic confinement device), faces two key challenges: (1) Plasma shape control -- 10-20 magnetic coils must be precisely coordinated in real time to maintain desired plasma configuration; (2) Disruption avoidance -- plasma instabilities can suddenly release stored energy onto the reactor wall, causing significant damage. AI addresses both: deep RL learns control policies by interacting with plasma simulators, while supervised deep learning predicts disruptions from multi-modal diagnostic signals. ## Detailed Analysis DeepMind's approach (2022): A deep RL agent was trained in a simulator to control 19 magnetic coils of the TCV tokamak. The policy network mapped sensor readings to coil currents, trained using a reward function penalizing deviations from the desired plasma shape. The agent discovered novel plasma configurations including the "droplet" shape (better heat exhaust) and "snowflake" divertor (spreads heat load). Crucially, the controller transferred successfully to the physical tokamak with minimal fine-tuning. Disruption prediction: Current tokamaks experience disruptions at rates of 10-30% of discharges, with ITER requiring <1% to avoid damage. ML models ingest multi-modal time-series data and predict disruptions 30-100ms in advance. DisruptionBench provides standardized train/test splits across multiple tokamaks, revealing that within-device prediction works well (F1>0.9) while cross-device generalization remains poor. The US Department of Energy (2025) demonstrated using deep RL to prevent tearing mode instabilities in real time. Beyond control: AI is also applied to accelerate fusion materials discovery, optimize stellarator coil designs, and analyze petabytes of plasma diagnostic data. ## Further Reading - DeepMind Blog: Accelerating fusion science through learned plasma control - TAE Technologies: Norman fusion device AI control - Commonwealth Fusion Systems / SPARC: AI for HTS magnet tokamak operations