Research map / Probabilistic and scientific modeling / Differential equations and PDEs
PARCv2: Physics-aware Recurrent Convolutional Neural Networks for Spatiotemporal Dynamics Modeling
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.
Extends the physics-aware recurrent convolution (PARC) differentiator-integrator with advection/diffusion/reaction terms and a hybrid numerical-plus-learned integrator (PARCv2), evaluated on Burgers, Navier-Stokes cylinder flow, and shock-induced energetic material hotspots.
Problem
Modeling unsteady, fast transient, and advection-dominated spatiotemporal field evolution where PDEs are unknown or hard to regularize with physics-informed losses, with sharp gradients and moving interfaces.
Approach
Adds explicit spatial gradient and Laplacian (central finite difference) operators into the PARC differentiator and a hybrid integrator that combines a conventional numerical integral (Heun or RK4) with a CNN-learned higher-order correction, trained in two stages without a physics-informed loss.
Key results
- PARCv2 achieves Navier-Stokes velocity RMSE of 0.0727 m/s versus 0.2094-0.2336 for baselines (about 1/4 of other methods).
- On energetic materials, PARCv2 temperature RMSE is 229.52 K versus about 249 K for PARC (numerical) and FNO.
- On Burgers, PARCv2 velocity RMSE is 0.0129 cm/s, roughly half of the other models (0.0236 to 0.0588), though PARC with numerical integration is slightly lower at 0.0074.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PARCv2 with hybrid integration instead of PARC with numerical integration | PARC differentiator-integrator · Burgers | RMSE_u (cm/s): 0.0074 -> 0.0129 | Hurt |
| use pure data-driven integration instead of numerical integration (ablation) | PARC · Burgers | RMSE_u (cm/s): 0.0074 -> 0.0236 | Hurt |
| use PARCv2 with hybrid integration instead of PARC with numerical integration | PARC differentiator-integrator · Navier-Stokes | RMSE_u (m/s): 0.2336 -> 0.0727 | Helped |
| use pure data-driven integration instead of numerical integration (ablation) | PARC · Navier-Stokes | RMSE_u (m/s): 0.2336 -> 0.3089 | Hurt |
| add data-driven higher-order correction to the numerical integrator | PARCv2 compared with PARC (numerical int.) · Navier-Stokes | RMSE_u (m/s): 0.2336 -> 0.0727 | Helped |
| add spatial gradient terms | PARCv2 compared with PARC · Energetic Materials | ε_A^hs (μm²): 0.0253 -> 0.0060 | Helped |
| use PARCv2 with hybrid integration instead of PARC with numerical integration | PARC differentiator-integrator · Energetic Materials | RMSE_P (GPa): 1.491 -> 1.634 | Hurt |
| use pure data-driven integration instead of numerical integration (ablation) | PARC · Energetic Materials | RMSE_P (GPa): 1.491 -> 4.111 | Hurt |
| use PARCv2 instead of the strongest baseline | PARCv2 · Navier-Stokes | ε_div (1/s): 0.0565 -> 0.3738 | Hurt |
| use PARCv2 instead of the strongest baseline | PARCv2 · Burgers | RMSE_u (cm/s): 0.0074 -> 0.0129 | Hurt |
Surprising findings
- PARCv2 violates the divergence-free condition substantially on Navier-Stokes (divergence error 0.3738 vs. 0.0398 for DNS), despite better RMSE.
- Two-stage training was empirically found to outperform end-to-end training, without a concrete explanation.
Limitations
- PARCv2 does not enforce physics residuals and shows higher PDE residual and divergence error than physics-informed baselines on Navier-Stokes.
- PARCv2 is computationally heavier than other PIML baselines on some tasks, e.g. 0.820 s on EM.
- The choice of numerical integrator (Heun vs RK4) is problem-dependent and chosen by hand.
Experimental setup
Datasets: Burgers equation simulations (implicit backward Euler, open-source code) Navier-Stokes cylinder flow (ANSYS Fluent) Energetic materials DNS (SCMITAR 3D)
Benchmarks: 2D Burgers equation Navier-Stokes cylinder flow energetic materials pore collapse
Metrics: RMSE PDE residual divergence error hotspot temperature and area errors
Baselines: PARC (numerical integrator) PARC (data-driven integrator) FNO PhyCRNet PIFNO
Builds on: PARC PhyCRNet Fourier Neural Operator FINN
Closest papers by meaning
- CALM-PDE: Continuous and Adaptive Convolutions for Latent Space Modeling of Time-dependent PDEs (NeurIPS 2025 · similarity 0.88)
- PINP: Physics-Informed Neural Predictor with latent estimation of fluid flows (ICLR 2025 · similarity 0.87)
- Sub-Sequential Physics-Informed Learning with State Space Model (ICML 2025 · similarity 0.87)
- Space and time continuous physics simulation from partial observations (ICLR 2024 · similarity 0.87)
- MultiPDENet: PDE-embedded Learning with Multi-time-stepping for Accelerated Flow Simulation (ICML 2025 · similarity 0.86)
- PINNacle: A Comprehensive Benchmark of Physics-Informed Neural Networks for Solving PDEs (NeurIPS 2024 · similarity 0.86)
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.