FUSE: Fast Unified Simulation and Estimation for PDEs
Authors: Levi Lingsch, Dana Grund, Siddhartha Mishra, Georgios Kissas
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments show that our implementation of FUSE overcomes struggles of baselines (c VANO, GAROM, In VAErt) and an ablation (U-Net) on two complex and realistic PDE examples, pulse wave propagation (PWP) in the human arterial network and an atmospheric cold bubble (ACB). The experiments exemplify its ability for accurate and fast surrogate modeling, parameter inference, out-of-distribution generalization, and the flexibility to handle different levels of input information. |
| Researcher Affiliation | Academia | Levi E. Lingsch Seminar for Applied Mathematics & AI Center, ETH Zurich levi.lingsch@ai.ethz.ch Dana Grund Institute for Atmospheric and Climate Science, ETH Zurich dana.grund@ethz.ch Siddhartha Mishra Seminar for Applied Mathematics & AI Center, ETH Zurich siddhartha.mishra@ethz.ch Georgios Kissas AI Center ETH Zurich gkissas@ai.ethz.ch |
| Pseudocode | No | The paper provides detailed mathematical formulations but does not include any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | Furthermore, we provide all code, data, and models (including baselines) required to reproduce the experiments outlined in this paper. |
| Open Datasets | Yes | Pulse Wave Propagation (PWP) in the Human Body ... through a reduced order PDE model in the dataset published by [7]. [7] P. H. Charlton, J. Mariscal Harana, S. Vennin, Y. Li, P. Chowienczyk, and J. Alastruey. Modeling arterial pulse waves in healthy aging: a database for in silico evaluation of hemodynamics and pulse wave indexes. American Journal of Physiology-Heart and Circulatory Physiology, 317 (5):H1062 H1085, 2019. doi:10.1152/ajpheart.00218.2019. |
| Dataset Splits | Yes | Table 3: Data details. Ntrain, Nval, and Ntrain provide the number of training, validation, and test samples. Experiment Ntrain Nval Ntest m Channels Sample Points Batch Size PWP 4000 128 128 32 52 487 64 ACB 8000 1000 1000 6 20 181 64 |
| Hardware Specification | Yes | All experiments were performed on an Nvidia Ge Forse RTX 3090 GPU with 24GB of memory. |
| Software Dependencies | No | The paper mentions several software components and models (e.g., FNO, FMPE, PyCLES), but does not specify their version numbers or other specific software dependencies required for reproducibility. |
| Experiment Setup | Yes | For each experiment, we performed a hyperparameter sweep when possible to select the learning rate, scheduler rate, weight decay, and network size. Due to the large number of hyperparameters, we selected 64 random combinations of hyperparameters and trained each model for 500 epochs. The final size of each model, in number of network parameters, is provided in Table 4. The Batch Size is 64 for both PWP and ACB experiments (Table 3). |