Research map / Probabilistic and scientific modeling / Molecules and proteins
CoilStellaration: A Dataset and Benchmark for Engineering-Aware Stellarator Coilset Generation
Jan-Hendrik Ewers, Santiago Cadena, Nicolò Foppiani, Andrea Merlo
Read the paperCheck your idea against 137,217 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.
Introduces CoilStellaration, a dataset and benchmark for engineering-aware stellarator coil set generation, aimed at letting machine learning methods design electromagnetic coils for fusion devices.
Problem
Designing electromagnetic coils that produce a stellarator's confining magnetic field is the key factor in whether the device can be practically engineered and built cost-effectively, and existing approaches lack a standard engineering-aware dataset and benchmark for coil set generation.
Approach
A dataset and benchmark for stellarator coil set generation that incorporates engineering considerations, so that coil designs can be evaluated for practical manufacturability and cost, not only magnetic field quality.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| ensemble five networks instead of a single model (ablation) | PerCoilMLP · variable-shape track | mean score S: 0.082 -> 0.137 | Improved |
| ensemble five networks instead of a single model (ablation) | ResMLP · fixed-shape track | mean score S: 0.076 -> 0.133 | Improved |
| use residual backbone instead of plain MLP (ablation) | MLP · fixed-shape track | strict feasibility: 2.6 -> 3.8 (single); 2.7 -> 4.4 (ensemble) | Improved |
| use residual backbone instead of plain MLP (ablation) | MLP · fixed-shape track | mean score S: 0.067 -> 0.076 | No clear effect |
| share one model across topologies (variable-shape) instead of fixed-shape specialist (ablation) | EnsemblePerCoilMLP · fixed-shape subset | mean score S: 0.133 -> 0.18 | Improved |
| tighter field-error requirement tier (ablation) | all baselines · requirement tier | strict feasibility: loose 7-12% -> tight 0% | Hurt |
| initialize DESC optimization with learned coilset instead of REGCOIL (ablation) | DESC · tight-tier instances, 500 iterations | strict feasibility: 18% -> 27% | Improved |
| initialize DESC optimization with learned coilset instead of REGCOIL (ablation) | DESC · tight-tier instances, 500 iterations | mean score S: 0.83 (REGCOIL) vs 0.76 (learned) | Hurt |
| initialize DESC optimization with learned coilset instead of REGCOIL (ablation) | DESC · tight-tier instances, early iterations (up to 200) | mean score S | Improved |
| initialize DESC optimization with learned coilset instead of REGCOIL (ablation) | DESC · tight-tier instances, 1000 iterations | strict feasibility: 26% -> 25% | No clear effect |
| learned start versus REGCOIL start on coil-coil spacing (ablation) | DESC · tight-tier instances, 1 iteration | coil-coil spacing requirement met: 92% -> 38% | Hurt |
| learned start versus REGCOIL start on coil-plasma spacing (ablation) | DESC · tight-tier instances, 1 iteration | coil-plasma spacing requirement met: 10% -> 58% | Improved |
Experimental setup
Benchmarks: CoilStellaration
Closest papers by meaning
- ConStellaration: A dataset of QI-like stellarator plasma boundaries and optimization benchmarks (NeurIPS 2025 · similarity 0.94)
- GyroSwin: 5D Surrogates for Gyrokinetic Plasma Turbulence Simulations (NeurIPS 2025 · similarity 0.81)
- Multi-Timescale Dynamics Model Bayesian Optimization for Plasma Stabilization in Tokamaks (ICML 2025 · similarity 0.78)
- The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning (NeurIPS 2024 · similarity 0.77)
- Diff-PIC: Revolutionizing Particle-In-Cell Nuclear Fusion Simulation with Diffusion Models (ICLR 2025 · similarity 0.76)
- EngiBench: A Framework for Data-Driven Engineering Design Research (NeurIPS 2025 · similarity 0.76)
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.
