Research map / Graph learning / Graph algorithms and clustering
Learning, Solving and Optimizing PDEs with TensorGalerkin: an efficient high-performance Galerkin assembly algorithm
Shizheng Wen, Mingyuan chi, Tianwei Yu, Ben Moseley, Mike Yan Michelis, Pu Ren, Hao Sun, Siddhartha Mishra
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace FEniCS (CPU) with TensorMesh (CPU) on mixed-BC Poisson circle domain | TensorMesh · Mixed-BC Poisson circle (6K nodes, CPU) | end-to-end time (ms): 7{,}000 -> 133 (~52x) | Improved |
| replace FEniCS (CPU) with TensorMesh (CPU) on mixed-BC Poisson boomerang domain | TensorMesh · Mixed-BC Poisson boomerang (14.8K nodes, CPU) | end-to-end time (ms): 5{,}600 -> 317 (~18x) | Improved |
| use TensorMesh instead of JAX-FEM for topology optimization optimization loop | TensorOpt · 2D cantilever beam topology optimization | time (s): 28.51 s -> 7.77 s (3.7x) | Improved |
| use TensorMesh instead of JAX-FEM for topology optimization setup | TensorOpt · 2D cantilever beam topology optimization | setup time (s): 2.62 s -> 0.58 s (4.5x) | Improved |
| use analytic-shape-gradient Galerkin loss (TensorPils) instead of Deep Ritz | Neural PDE solver · Poisson checkerboard K=8 | relative L2 error (%): 10.60 -> 10.05 | Improved |
| use TensorPils instead of PINN strong-form loss | Neural PDE solver · Poisson checkerboard K=8 | relative L2 error (%): 34.77 -> 10.05 | Improved |
| use TensorPils instead of PINN for training throughput (LBFGS) | Neural PDE solver · Poisson checkerboard | training speed (it/s): 1.0 -> 99.7 | Improved |
| use physics-informed TensorPils instead of data-driven training with same GNN (OOD wave) | GNN operator · Wave operator learning OOD | relative L2 error: 0.230 -> 0.090 | Improved |
| use TensorPils instead of data-driven training in-distribution (wave) | GNN operator · Wave operator learning ID | relative L2 error: 0.089 -> 0.085 | No clear effect |
| use TensorPils instead of PI-DeepONet (AC OOD) | operator learning · Allen-Cahn operator learning OOD | relative L2 error: 8.536 -> 0.083 | Improved |
| use TensorPils loss instead of PINN loss on unstructured meshes (forward pass) | loss evaluation · unstructured triangular meshes | forward loss time: roughly 4--6$\times$ faster | Improved |
| use TensorMesh GPU instead of second-best FEniCS for Poisson solving | TensorMesh · Poisson, up to 1e6 DoFs | runtime: 3 times speedup | Improved |
Closest papers by meaning
- Toward Efficient Kernel-Based Solvers for Nonlinear PDEs (ICML 2025 · similarity 0.87)
- Physics-informed Reduced Order Modeling of Time-dependent PDEs via Differentiable Solvers (NeurIPS 2025 · similarity 0.86)
- Holistic Physics Solver: Learning PDEs in a Unified Spectral-Physical Space (ICML 2025 · similarity 0.86)
- MC^2: Monte Carlo Correction for Fast Elliptic PDE Solving (NeurIPS 2026 · similarity 0.86)
- Mosaic: A Benchmark Suite for Differentiable Physics Solvers (NeurIPS 2026 · similarity 0.86)
- PhyMPGN: Physics-encoded Message Passing Graph Network for spatiotemporal PDE systems (ICLR 2025 · 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.
