physics.plasm-phMar 16, 2026
SaveA proof-of-concept for automated AI-driven stellarator coil optimization with in-the-loop finite-element calculations
Organizations: Courant Institute of Mathematical Sciences, New York University, New York, NY 10012, USA · Max-Planck-Institut für Plasmaphysik, 85748 Garching, Germany
Abstract
Finding feasible coils for stellarator fusion devices is a critical challenge of realizing this concept for future power plants. Years of research work can be put into the design of even a single reactor-scale stellarator design. To rapidly speed up and automate the workflow of designing stellarator coils, we have designed an end-to-end ``runner'' for performing stellarator coil optimization. The entirety of pre and post-processing steps have been automated; the user specifies only a few basic input parameters, and final coil solutions are updated on an open-source leaderboard. Two policies are available for performing non-stop automated coil optimizations through a genetic algorithm or a context-aware LLM. Lastly, we construct a novel in-the-loop optimization of Von Mises stresses in the coils, opening up important future capabilities for in-the-loop finite-element calculations.
Figures & tables
| Component | Motivation | Outcome |
| Automated coil initialization | Manual coil initialization is usually required for different plasma surfaces. | Circular coils are initialized without coil-coil or coil-surface intersections, and modular coils always interlink the plasma. |
| Automated coil optimization | Manual weight tuning is expensive and very different weights are needed for different plasmas. | Objective weights are no longer user-defined; they are auto-tuned with a penalty method. All SIMSOPT functionality is available. |
| Genetic or LLM-driven policy | Learning the optimal hyperparameters such as thresholds, convergence tolerances, number of iterations, “schedules”, etc. is very challenging. | Genetic algorithm or context-aware LLM makes the decisions about which runs to submit every batch. Can keep a running log of justifications. |
| Automated post-processing | Post-processing requires installing a host of complex simulation codes and substantial training of researchers to accurately use the codes. | Users can run new cases and plasma surfaces on cloud devices (through the continuous integration) where all post-processing is auto-installed and auto-run |
| Automated documentation & leaderboard | Historically, it has been challenging to make precise comparisons between stellarator coil solutions. | Runs receive an overall composite score by a weighted sum of coil objectives. Runs are added in real-time to an open-source, online ReadTheDocs leaderboard. |
| Reproducibility | Historically, it has been challenging to make precise comparisons between stellarator coil solutions. | Plasma surfaces and other parameters are standardized and cannot be altered. Full metadata, all final solutions, and all metrics are saved in a database. |
Table 1: Workflow components for the automated code runner.
Figure 1: Automated code runner interactions with external codes and the genetic or LLM-proposer.
Figure 2: Result of post-processing routines. (a) Shape gradients evaluated along final coils; (b) finite-build and Von Mises stresses; (c) coil sensitivity analysis; (d) sample Poincaré plot; (e) sample Boozer plot at ; (f) loss fraction plot; (g) two-term quasisymmetry profile.
Figure 3: Left: Final coilset obtained with an order-four Landreman-Paul QA coil optimization with the default parameters. Right: Same exact optimization except there is also a penalty on volume-averaged Von Mises stress above 0.01 GPa. Note that the field accuracy decreases but the maximum Von Mises stress is reduced by about an order of magnitude.
Figure 4: Top: scatter plot showing the maximum curvature versus field accuracy achieved in the 443 optimized coilsets. Note that the values plotted here correspond to the Landreman-Paul QA equilibrium scaled to 1 m major radius. The total coil length per half-field period is plotted as well. A front forms at around due to the imposed threshold on the squared flux. A cluster is visible at high error magnitudes, due to the exploration policy. Bottom: one of the notable configurations from this preliminary implementation of the genetic algorithm, a 3-coil per half-field period with relatively low curvature, short coils, low field errors and high coil-to-surface distance.
| Property | 3-coil QA (policy) | 3-coil QA [ 22 ] |
| Total coil length (m) | 182.0 | |
| Max curvature, (m -1 ) | 0.53 | |
| Max mean-squared , (m -2 ) | 0.05 | |
| Min coil–coil dist., (m) | ||
| Min coil–surface dist., (m) | ||
Table 2: Reactor-scale coil properties for the 3-coil per hfp QA coilset. Numbers are reported for the unique set of coils only. ∗ For fair comparison we scaled the numbers from [ 22 ] using 496 turns and assume this number for each coil. Critical current density was not used in the previous work for determining the number of turns.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Parameter | Default | Description |
| width , height | 0.35 m | Cross-section dimensions |
| E | 100 GPa | Young’s modulus |
| nu | 0.3 | Poisson ratio |
| stress_metric | Scalar metric | |
| fd_step | Finite-difference step | |
| eval_interval | 1 | Evaluate every eval_interval iterations |
Table 4: The key implementation parameters for the in-the-loop optimization of Von Mises stresses. We emphasize that most of these parameters can be set to the default values.
Figure 5: Convergence plots for the default settings showing (slow) convergence as the mesh resolution increases. Note the y-axes; the average displacements and average stresses do not change much in absolute terms from low to high resolution.
Figure 6: Log-scale of the Von Mises stresses for the highest resolution solution in the convergence test.