Research map / Image generation / Fast diffusion sampling
Diff-PIC: Revolutionizing Particle-In-Cell Nuclear Fusion Simulation with Diffusion Models
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.
Diff-PIC distills PIC laser-plasma interaction electric-field snapshots into a physically-informed conditional diffusion model, made one-step via rectified flow, for fast surrogate generation.
Problem
Particle-in-Cell (PIC) simulations of laser-plasma interaction in inertial confinement fusion are computationally prohibitive, creating a bottleneck for fusion research.
Approach
Conditional diffusion model with a Physically-Informed Parameter Encoder (positional encoding plus polynomial basis encoding of Te, Ti, I and snapshot time) and rectified flow reflow to collapse sampling into a single step, generating PIC-like electric field snapshots at arbitrary times.
Key results
- Diff-PIC achieves 1.62e4x speedup over 100 ps PIC simulation on RTX 4090 GPU.
- Interpolation test MAE 1.68e-2 / RMSE 2.29e-2 / FID 1.62 on E1, outperforming GAN-PIC and NF-PIC.
- Average reduction of 59.21% MAE, 57.15% RMSE, 39.46% FID versus two generative baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Diff-PIC instead of the strongest baseline, NF-PIC | Electric-field generation · Interpolation test E1 and E2 | MAE: 4.61e-2 -> 1.68e-2; 1.86e-2 -> 8.26e-3 | Helped |
| Use Diff-PIC instead of the strongest baseline, NF-PIC | Electric-field generation · Extrapolation E1 and E2, 10% and 20% extensions | MAE: 4.74e-2 -> 1.83e-2; 5.41e-2 -> 2.18e-2; 1.90e-2 -> 9.47e-3; 2.63e-2 -> 1.13e-2 | Helped |
| Use Diff-PIC-GPU instead of PIC | 100 ps simulation · PIC-100 ps | speedup: 1.00$\times$ -> 1.62e4$\times$ | Helped |
| Use Diff-PIC-GPU instead of PIC | 100 ps simulation · PIC-100 ps | Energy reduction: 1.00$\times$ -> 1.01e4$\times$ | Helped |
| Add a Physically-Informed Parameter Encoder | Diff-PIC · LPI parameter generalization | generalizability | Helped |
| Add Rectified Flow Acceleration for one-step generation | Diff-PIC · Electric-field generation | generation efficiency | Helped |
Limitations
- Extrapolation evaluated only up to 10% and 20% beyond training ranges.
- Only electric field snapshots in 2D are modeled; other physical quantities not covered.
- Speedup comparison uses a 100 ps reference and differing hardware between PIC and generative models.
Experimental setup
Datasets: OSIRIS PIC simulation data (6,615 simulations, 80 snapshots each)
Benchmarks: Diff-PIC interpolation/extrapolation split of OSIRIS simulations
Metrics: MAE RMSE FID speedup energy consumption reduction
Baselines: GAN-PIC NF-PIC PIC simulation (OSIRIS)
Builds on: diffusion models rectified flow positional encoding U-Net
Closest papers by meaning
- GyroSwin: 5D Surrogates for Gyrokinetic Plasma Turbulence Simulations (NeurIPS 2025 · similarity 0.82)
- Multi-Timescale Dynamics Model Bayesian Optimization for Plasma Stabilization in Tokamaks (ICML 2025 · similarity 0.82)
- M2PDE: Compositional Generative Multiphysics and Multi-component PDE Simulation (ICML 2025 · similarity 0.81)
- Particle Guidance: non-I.I.D. Diverse Sampling with Diffusion Models (ICLR 2024 · similarity 0.80)
- ConStellaration: A dataset of QI-like stellarator plasma boundaries and optimization benchmarks (NeurIPS 2025 · similarity 0.80)
- The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning (NeurIPS 2024 · similarity 0.79)
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.