Research map / Probabilistic and scientific modeling / Dynamical systems
GyroSwin: 5D Surrogates for Gyrokinetic Plasma Turbulence Simulations
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.
Introduces GyroSwin, a multitask hierarchical Swin-based 5D neural surrogate that evolves the 5D gyrokinetic distribution function and predicts 3D potentials and heat flux, scaling to ~1B parameters.
Problem
Modeling plasma turbulence requires nonlinear 5D gyrokinetic simulations, which are computationally prohibitive; reduced quasilinear models omit nonlinear effects such as zonal flows.
Approach
Extends hierarchical Swin Transformer U-Nets to 5D with shifted 5D window attention, latent 5D<->3D cross-attention mixing and velocity-space integrator modules, multitask heads for f, potential and heat flux, and channelwise mode separation that isolates the zonal flow (k_y=0) mode as extra channels.
Key results
- GyroSwin-Large (998M params, 241 simulations) reaches ID/OOD Q(k_y) correlation 0.87/0.91, versus 0.51/0.58 for QL.
- On the 48-simulation set, GyroSwin reduces heat flux RMSE to 47.35 (ID) vs 89.48 for QL, and 16.37 for the Small model trained on 241 simulations.
- Scaling laws fit a joint power law in model and data size; the fitted exponents are alpha=0.46 and beta=0.37, and the model is reported as about three orders of magnitude faster than GKW in FLOPs (756 vs 4200 GFLOPs).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GyroSwin instead of ViT for autoregressive 5D turbulence prediction | 5D neural surrogates · 5D turbulence modelling, ID and OOD | correlation time: 16.83 $\pm$ 1.49 -> 17.67 $\pm$ 2.16; 19.20 $\pm$ 1.36 -> 19.40 $\pm$ 1.46 | Helped |
| Scale GyroSwin training from 48 to 241 simulations | GyroSwin · Average heat flux ID and OOD | RMSE: 47.35 $\pm$ 6.20 -> 16.37 $\pm$ 3.68; 59.15 $\pm$ 15.62 -> 16.73 $\pm$ 5.58 | Helped |
| Scale GyroSwin training from 48 to 241 simulations | GyroSwin · Flux spectrum Q(ky) ID and OOD | Pearson correlation: 0.59 $\pm$ 0.06 -> 0.87 $\pm$ 0.11; 0.69 $\pm$ 0.06 -> 0.91 $\pm$ 0.08 | Helped |
| Increase model size from 239M to 893M parameters at full data | GyroSwin · Validation | loss: 12% | Helped |
| Use DiT-style conditioning instead of FiLM (ablation) | GyroSwin · Gyrokinetic prediction | performance | No clear effect |
| Center the nonlinear flux vector when fitting the QuaLiKiz normalization (ablation) | QuaLiKiz saturation rule · Nonlinear flux prediction | error | Helped |
| Use GyroSwin instead of QL | Reduced numerical modelling · Flux spectrum Q(ky), ID and OOD | Pearson correlation: 0.51 $\pm$ 0.4 -> 0.59 $\pm$ 0.06; 0.58 $\pm$ 0.33 -> 0.69 $\pm$ 0.06 | Helped |
| Scale GyroSwin from 90.2M to 998.3M parameters | GyroSwin · Flux spectrum Q(ky), ID and OOD | Pearson correlation: 0.82 $\pm$ 0.10 -> 0.87 $\pm$ 0.11; 0.79 $\pm$ 0.09 -> 0.91 $\pm$ 0.08 | Mixed |
| Scale GyroSwin from small to medium model at 241 simulations | GyroSwin · Average heat flux ID and OOD | RMSE: 16.37 $\pm$ 3.68 -> 20.41 $\pm$ 3.86; 16.73 $\pm$ 5.58 -> 24.96 $\pm$ 5.63 | Mixed |
Surprising findings
- Most neural surrogates converge to nearly identical flux errors (~120) because predicted turbulent transport decays to zero along rollouts.
- Model size and data size must be scaled together; small models cannot exploit additional data and vice versa.
Limitations
- Does not model the distributional/chaotic nature of turbulence; error accumulates over rollouts despite stability beyond 100 steps.
- Neglects the linear phase of simulations and only considers the adiabatic electron approximation.
- Only ion temperature gradient turbulence (cyclone base case) and four operating parameters are covered.
Experimental setup
Datasets: GKW nonlinear gyrokinetic simulations (255 simulations, Cyclone base case, adiabatic electrons)
Benchmarks: in-distribution and out-of-distribution gyrokinetic test sets (6 ID, 5 OOD simulations)
Metrics: RMSE of time-averaged heat flux correlation time of 5D rollouts (Pearson threshold 0.1) Pearson correlation of Q(k_y) and W(k_y) forward time, memory, parameter count
Baselines: QuaLiKiz saturation rule (QL) Gaussian process regression MLP FNO PointNet Transolver ViT
Builds on: Swin Transformer Swin UNet Perceiver FiLM SwinV2
Closest papers by meaning
- Multi-Timescale Dynamics Model Bayesian Optimization for Plasma Stabilization in Tokamaks (ICML 2025 · similarity 0.84)
- Calibrated Physics-Informed Uncertainty Quantification (ICML 2025 · similarity 0.84)
- M2PDE: Compositional Generative Multiphysics and Multi-component PDE Simulation (ICML 2025 · similarity 0.83)
- Metamizer: A Versatile Neural Optimizer for Fast and Accurate Physics Simulations (ICLR 2025 · similarity 0.83)
- Universal Physics Transformers: A Framework For Efficiently Scaling Neural Operators (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.