DeepLag: Discovering Deep Lagrangian Dynamics for Intuitive Fluid Prediction
Authors: Qilong Ma, Haixu Wu, Lanxiang Xing, Shangchen Miao, Mingsheng Long
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimentally, Deep Lag excels in three challenging fluid prediction tasks covering 2D and 3D, simulated and real-world fluids. |
| Researcher Affiliation | Academia | Qilong Ma , Haixu Wu , Lanxiang Xing, Shangchen Miao, Mingsheng Long School of Software, BNRist, Tsinghua University, China {mql22,wuhx23,xlx22,msc21}@mails.tsinghua.edu.cn, mingsheng@tsinghua.edu.cn |
| Pseudocode | No | No explicit "Pseudocode" or "Algorithm" section found. The description is in text and mathematical formulas. |
| Open Source Code | Yes | Code is available at this repository: https://github.com/thuml/Deep Lag. |
| Open Datasets | Yes | First, we downloaded daily sea reanalysis data [5] from 2011 to 2020 provided by the ECMWF. [5] CMEMS and MDS. Global ocean physics reanalysis. DOI: 10.48670/moi-00021 (Accessed on 23 September 2023), 2023. |
| Dataset Splits | No | 2000 sequences with spatial resolution of 128 × 128 are generated for training and 200 new sequences are used for the test. |
| Hardware Specification | Yes | All experiments are implemented in Py Torch [24] and conducted on a single NVIDIA A100 GPU. |
| Software Dependencies | No | All experiments are implemented in Py Torch [24] |
| Experiment Setup | Yes | Deep Lag is trained with relative L2 as the loss function on all benchmarks. We use the Adam [18] optimizer with an initial learning rate of 5 × 10−4 and Step LR learning rate scheduler. The batch size is set to 5, and the training process is stopped after 100 epochs. (From main text) And Table 7: Model Designs Hyperparameters Values Number of observation steps P 10 Number of scales L 4 Eulerian-Lagrangian Sample Points at each scale {M1, ..., ML} {512, 128, 32, 8} Recurrent Network Downsample Ratio r = |Dl+1| / |Dl| 0.5 Channels of each scale {C1, ..., CL} {64, 128, 256, 256} Paddings for Ocean Current dataset (12, 20) Eu Lag Block Heads in Cross-Attention 8 Channels per head in Cross-Attention 64 (From Table 7, Appendix A.1) |