Research map / Probabilistic and scientific modeling / Bayesian optimization
Multi-Timescale Dynamics Model Bayesian Optimization for Plasma Stabilization in Tokamaks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DynaBO, a multi-timescale Bayesian optimization method that pairs a recurrent probabilistic neural network dynamics model (as GP prior mean) with a GP over marginal statistics, and validates it offline and on 8 live DIII-D tokamak shots for avoiding tearing instabilities.
Problem
Finding ECH heating profiles that avoid 2-1 tearing mode instabilities in high-performance tokamak plasmas, where data is scarce, dynamics shift between experiments, and only a handful of experiments are affordable.
Approach
Combine a high-frequency RPNN dynamics model, used as the prior mean of a low-dimensional contextual GP over target normalized pressure and Gaussian ECH parameters, and run UCB-based contextual BO that updates the GP between shots using measured (not requested) inputs.
Key results
- In live DIII-D experiments, DynaBO avoided tearing instabilities in 4 of 8 shots (50%), versus a 77% historical tearing rate for similar configurations.
- Mean stability time was 3339 ms with DynaBO versus 2424 ms historically, a 914 ms improvement (about 38%; the paper reports a 117% improvement in the abstract by avoidance rate).
- In offline simulation, DynaBO and DynaBO with time-dependent kernel achieved the lowest cumulative regret across most kernels, outperforming RPNN-only and vanilla GP baselines; linear kernel was the exception.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DynaBO to choose ECH profiles instead of historical manual selection | DIII-D high q_min experiments · DIII-D high q_min live experiments | mean stability time: 2424 milliseconds -> 3339 milliseconds; 914 ms improvement | Helped |
| Use DynaBO to choose ECH profiles instead of historical manual selection | DIII-D high q_min experiments · DIII-D high q_min live experiments | tearing instability avoidance: 4/8; historical ... 77% | Helped |
| Add a time-dependent kernel to the GP (ablation) | DynaBO · Offline historical data | cumulative regret | No clear effect |
| Replace the dynamic-model prior mean with a zero-mean prior (ablation) | DynaBO · Offline historical data | cumulative regret | Hurt |
| Replace the dynamic-model prior mean with the mean of past data (ablation) | DynaBO · Offline historical data | cumulative regret | Hurt |
| Use RPNN predictions without GP updates (ablation) | DynaBO · Offline historical data | cumulative regret | Hurt |
| Use a linear kernel (ablation) | DynaBO · Offline historical data | cumulative regret | Hurt |
| Vary the acquisition function among UCB, Thompson Sampling, and Expected Improvement (ablation) | DynaBO · Offline historical data, 500 epochs | cumulative regret: UCB, Thompson Sampling, and EI have similar order cumulative regret | No clear effect |
| Add the dynamic-model prior mean to vanilla GP (ablation) | Vanilla GP · Offline historical data | exploration | Helped |
Surprising findings
- Adding time as a GP input gave only marginal improvement, which the authors attribute to reliability depending on factors not explained by a continuous function of time.
- Feeding tearing mode labels into the RPNN did not yield good results, attributed to spurious correlations and causality issues.
Limitations
- Only applicable to feedforward control; cannot adapt to real-time plasma changes such as MHD activity or impurity changes.
- Live experiment of only 8 shots on one configuration is too few for statistical significance, as the authors acknowledge.
- Method is mainly data-driven and does not incorporate physics knowledge such as the Rutherford equation.
Experimental setup
Datasets: DIII-D historical tokamak experiment data (2010-2019 for RPNN; 2012-2023 for GP)
Benchmarks: offline replay of DIII-D historical high q_min data (281 trajectories) live DIII-D experiments (8 shots)
Metrics: cumulative regret (time-to-tearing) tearing instability avoidance rate mean stability time (ms)
Baselines: RPNN-only policy vanilla GP with zero-mean prior vanilla GP with mean-value prior DynaBO with time-dependent kernel historical manual ECH experiments
Builds on: Bayesian optimization contextual Bayesian optimization upper confidence bound
Closest papers by meaning
- Calibrated Physics-Informed Uncertainty Quantification (ICML 2025 · similarity 0.84)
- GyroSwin: 5D Surrogates for Gyrokinetic Plasma Turbulence Simulations (NeurIPS 2025 · similarity 0.84)
- Parametric model reduction of mean-field and stochastic systems via higher-order action matching (NeurIPS 2024 · similarity 0.83)
- Stable Port-Hamiltonian Neural Networks (NeurIPS 2025 · similarity 0.82)
- Sample-efficient Bayesian Optimisation Using Known Invariances (NeurIPS 2024 · similarity 0.82)
- Diff-PIC: Revolutionizing Particle-In-Cell Nuclear Fusion Simulation with Diffusion Models (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.