Towards Universal Mesh Movement Networks

Authors: Mingrui Zhang, Chunyang Wang, Stephan C. Kramer, Joseph G Wallwork, Siyi Li, Jiancheng Liu, Xiang Chen, Matthew Piggott

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5 Experiments, We evaluate our method on advection and Navier-Stokes based examples, as well as a real-world tsunami simulation case. Our method out-performs existing learning-based mesh movement methods in terms of the benchmarks described above.
Researcher Affiliation Collaboration 1Imperial College London 2University of Cambridge 3Noah s Ark Lab, Huawei
Pseudocode Yes Algorithm 1 Mesh Movement Enhanced Simulation
Open Source Code Yes The code used to produce the results presented in this paper is publicly available at https:// github.com/mesh-adaptation/UM2N. (Release [49].)
Open Datasets No Aiming here to train universal mesh movement networks, we construct a PDE-independent training dataset D = {d = (ξ, mξ; V r, T r)}... To build the dataset, we randomly generate generic solution fields...
Dataset Splits No The training dataset consists of 600 randomly generated generic solution fields and original meshes with 463 and 513 vertices.
Hardware Specification Yes The training and experiments are performed on an Nvidia RTX 3090 GPU.,We tested the inference time of UM2N on two GPUs (Nvidia Ge Force RTX 3090 and Nvidia A100) and compared its performance to the MA method. ... The CPU used for testing was an 11th Gen Intel(R) Core(TM) i9-11900K @ 3.50GHz.
Software Dependencies No PDEs are solved using finite element methods with Firedrake [44]. Firedrake is written in Python, uses the Unified Form Language (UFL) [45] domain-specific language... Firedrake uses PETSc [46] for solving linear and nonlinear equations... Original meshes (i.e., meshes before adaptation) are generated by either Firedrake or Gmsh [47]... The conventional mesh movement strategy used to generate training data is implemented in Movement [40]...
Experiment Setup No The training dataset consists of 600 randomly generated generic solution fields and original meshes with 463 and 513 vertices. The model is trained using the Adam optimizer.