Bayesian optimization of poloidal field coil positions in tokamaks

Timothy Nunn, Kamran Pentland, Vignesh Gopakumar, John L. BuchananPhysics 2025 · DOI

TL;DR

This paper introduces a Bayesian Optimization (BO) framework to efficiently design the complex poloidal field (PF) coil system for tokamaks, crucial for fusion energy. By intelligently exploring the design space, BO significantly outperforms traditional methods, finding optimal coil configurations with fewer computationally expensive simulations.

What Problem Does This Paper Address?

Designing a tokamak, a leading concept for fusion energy, involves a complex trade-off between numerous physical and engineering objectives. Specifically, determining the optimal positions for the poloidal field (PF) coils is a critical challenge. These coils are essential for confining and shaping the plasma, but their placement must balance factors like minimizing manufacturing costs (coil volume) and maximizing plasma performance (connection length) while adhering to strict engineering constraints (e.g., avoiding intersections with other components, maintaining plasma stability). Traditional methods often struggle with this multi-objective, computationally intensive problem, requiring extensive simulations. This paper aims to address this by demonstrating how Bayesian Optimization can efficiently navigate this complex design space.

Methods

The core of this work is the application of Multi-Objective Bayesian Optimization (BO). BO is an iterative machine learning technique that uses probabilistic models, specifically Gaussian Processes (GPs), to build a "surrogate model" of the objective functions. This surrogate model allows the algorithm to predict the performance of a given design without running a full, expensive simulation.

Key methodological components include:

  • Gaussian Process Surrogate Model: GPs are used to model the relationship between PF coil positions (inputs) and the resulting objectives and constraints (outputs). They provide not only predictions but also uncertainty estimates, which are crucial for guiding the search.
  • Failure Region Classifier: Since the plasma equilibrium simulator can fail for certain coil configurations, a GP-based classifier is trained to predict these "failure regions." This prevents the BO from wasting computational resources on invalid designs.
  • Expected Hypervolume Improvement (EHVI) Acquisition Function: This function guides the BO by identifying the next most "informative" design to evaluate. EHVI balances exploring uncertain regions of the design space with exploiting regions that are likely to yield optimal solutions, aiming to maximize the hypervolume of the Pareto front.
  • Constrained Acquisition Function (ECHVI): To handle the numerous constraints, the EHVI is modified to incorporate a probability of feasibility, effectively prioritizing designs that are likely to satisfy all engineering requirements.

The framework is designed to be flexible, allowing for various objectives and constraints, and is machine-agnostic, meaning it can be applied to different tokamak designs.

Datasets

The primary "dataset" used in this work is generated by a plasma equilibrium simulator called FreeGS. This simulator takes PF coil configurations as input and outputs plasma equilibrium properties, which are then used to calculate the objectives and constraints.

  • STEP Baseline Design: The paper uses the initial PF coil setup and limiter geometry from the baseline STEP design as a reference point. This baseline configuration serves as a benchmark against which the optimized designs are compared.
  • Generated Data: The BO process iteratively generates data by evaluating FreeGS for proposed PF coil configurations. This data is used to train and update the GP surrogate models and the failure classifier.

No pre-existing public datasets were used for training; the data was generated specifically for this study.

Key Insights and Ideas

  • BO's Data Efficiency: The most significant insight is BO's remarkable data efficiency. It can identify high-quality, Pareto-optimal PF coil designs using a fraction of the computational budget required by traditional methods like Sobol sampling or genetic algorithms. This is crucial for fusion design, where high-fidelity simulations are extremely time-consuming.
  • Balancing Exploration and Exploitation: The use of GPs and acquisition functions effectively balances exploring the vast design space for novel solutions and exploiting promising regions to refine existing designs. This is key to finding truly optimal trade-offs.
  • Handling Failure Regions: The integration of a failure classifier is a critical innovation. It prevents the BO from wasting valuable simulation time on configurations that lead to simulation failures or physically invalid equilibria, significantly improving efficiency.
  • Pareto Front Identification: BO successfully identifies a Pareto front, representing a set of optimal trade-offs between competing objectives (e.g., coil volume vs. connection length). This provides designers with a range of high-performing, yet distinct, design options.
  • Generalizability: The framework's design emphasizes flexibility, suggesting its applicability to a broader range of fusion engineering design problems beyond just PF coils.

Key Results

  • Outperforms Traditional Methods: In experiments with a fixed computational budget, the BO framework achieved a significantly larger hypervolume of feasible solutions compared to Sobol sampling and the NSGA-II genetic algorithm. For example, with 128 evaluations, BO achieved a hypervolume of 1990.9, while the best genetic algorithm experiment (with 8x more samples) reached 1820.3.
  • Improved Designs: The BO identified PF coilsets that offered improvements over the baseline design. For instance, one optimized sample (Sample 110) reduced the volumetric sum of PF coils by 10.3% (from 86.5 m³ to 77.6 m³) while increasing the average connection length by 0.4% (from 96.5 m to 96.9 m). Another (Sample 71) achieved a higher connection length (71.8 m vs. 67.1 m) with a slightly larger coil volume.
  • High Feasibility Rate: As the BO progressed, the proportion of feasible samples increased dramatically. In the initial 64 Sobol samples, only 10% were feasible; after 64 BO iterations, 36% of the samples were feasible, with many being Pareto optimal.
  • Constraint Adherence: The optimized solutions successfully adhered to all engineering constraints, demonstrating the effectiveness of the constrained acquisition function.

Limitations and Future Work

  • Computational Cost of High-Fidelity Simulations: While BO is data-efficient, the underlying high-fidelity simulations (FreeGS) remain computationally expensive. The paper acknowledges that further improvements in simulation speed would accelerate the design process even more.
  • Dimensionality: The current framework can be adapted to higher-dimensional problems (e.g., including PF coil shaping), but this may require dimensionality reduction techniques.
  • Additional Objectives/Constraints: The authors suggest incorporating more objectives and constraints to further refine designs and improve performance (e.g., maximizing flux expansion for divertor performance).
  • Extension to Other Design Problems: The framework's generalizability is highlighted as a key future direction, suggesting its application to other complex fusion engineering challenges like magnetic sensor placement or stellarator coil design.
  • Simulator Improvements: The paper notes that an updated version of the FreeGS simulator (FreeGSNKE) can inherently handle some constraints, potentially simplifying the BO setup by removing the need for explicit constraint handling in the acquisition function.

Summary updated

Abstract

The tokamak is a world-leading concept for producing sustainable energy via magnetically confined nuclear fusion. Identifying where to position the magnets within a tokamak, specifically the poloidal field (PF) coils, is a design problem which requires balancing a number of competing economical, physical, and engineering objectives and constraints. In this paper, we show that multi-objective Bayesian optimization (BO), an iterative optimization technique utilizing probabilistic machine learning models, can effectively explore this complex design space and return several optimal PF coilsets. These solutions span the Pareto front, a subset of the objective space that optimally satisfies the specified objective functions. We outline an easy-to-use BO framework and demonstrate that it outperforms alternative optimization techniques while using significantly fewer computational resources. Our results show that BO is a promising technique for fusion design problems that rely on computationally demanding high-fidelity simulations.
BibTeX
@inproceedings{nunn2025bayesian,
  author = {Timothy Nunn and Kamran Pentland and Vignesh Gopakumar and John L. Buchanan},
  title = {{Bayesian optimization of poloidal field coil positions in tokamaks}},
  booktitle = {Physics},
  year = {2025},
  doi = {10.1063/5.0272085},
}
Figure