Physics-aligned field reconstruction with diffusion bridge
Authors: Zeyu Li, Hongkun Dou, Shen Fang, Wang Han, Yue Deng, Lijun Yang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the effectiveness of Pal SB through its application to three complex nonlinear systems: cylinder flow from Particle Image Velocimetry experiments, two-dimensional turbulence, and a reaction-diffusion system. The results reveal that Pal SB not only achieves higher accuracy but also exhibits enhanced compliance with physical constraints compared to existing methods. This highlights Pal SB s capability to generate high-quality representations of intricate physical interactions, showcasing its potential for advancing field reconstruction techniques. The source code can be found at https://github.com/lzy12301/Pal SB. |
| Researcher Affiliation | Academia | Zeyu Li, Hongkun Dou, Shen Fang, Wang Han, Yue Deng, Lijun Yang Beihang University EMAIL |
| Pseudocode | Yes | Algorithm 1 Pretraining of Pal SB Algorithm 2 Physics-aligned finetuning of Pal SB Algorithm 3 Sampling process of Pal SB |
| Open Source Code | Yes | The source code can be found at https://github.com/lzy12301/Pal SB. |
| Open Datasets | Yes | We use the high-fidelity dataset with spatial resolution of 256 256 published in (Shu et al., 2023), which contains 40 trajectories with 320 frames in each trajectory. The training and testing dataset is drawn from PDEBench Takamoto et al. (2022) with η = ζ = 1e 8, periodic boudaries and turbulent initial conditions, which contains 600 trajectories of v, ρ and p. |
| Dataset Splits | Yes | Cylinder flow measured by PIV. ...the former 70% of the trajectory is used for training and the rest is used for test. 2D forced turbulence. ...We use 90% of the trajectories as the training data and the rest 10% as the test data. Reaction-diffusion system. ...Two trajectories with 3000 frames that start with different initial conditions are used for training and testing, respectively. B.1 EXTENSION ON 3D DATASET. ...We train our model on the first 90% of the trajectories and test on the rest. |
| Hardware Specification | Yes | All the experiments conducted in this paper are running on a single GPU of Nvidia Ge Force RTX 3090 with Intel(R) Xeon(R) Gold 6226R CPU. |
| Software Dependencies | No | The platform is Ubuntu 20.04.3 LTS operation system with Python 3.9 environment. No specific library versions or solvers are mentioned. |
| Experiment Setup | Yes | The hyperparameters of the network architecture are listed in Table 6. The hyperparameters of FNO we used are listed in Table 7. We list the parameters for reproducing our experiments in Tables 8 and 9, including the training, finetuning and sampling process. Table 8: Parameters for training (Optimizer, Learning rate, Batch size, Number of iterations, Sampling steps T, Sampling step size, Backpropagation steps B, etc.) Table 9: Loss weights for finetuning (γphys, γreg) |