CL-DiffPhyCon: Closed-loop Diffusion Control of Complex Physical Systems
Authors: Long Wei, Haodong Feng, Yuchen Yang, Ruiqi Feng, Peiyan Hu, Xiang Zheng, Tao Zhang, Dixia Fan, Tailin Wu
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate CL-Diff Phy Con on two tasks: 1D Burgers equation control and 2D incompressible fluid control. The results demonstrate that CL-Diff Phy Con achieves superior control performance with significant improvements in sampling efficiency. The code can be found at https://github.com/AI4Science-Westlake U/CL_Diff Phy Con. |
| Researcher Affiliation | Academia | 1Department of Artificial Intelligence, Westlake University, 2School of Statistics and Data Science, Nankai University, 3Academy of Mathematics and Systems Science, Chinese Academy of Sciences, 4School of Future Technology, South China University of Technology |
| Pseudocode | Yes | Algorithm 1 Closed-loop Control of CL-Diff Phy Con |
| Open Source Code | Yes | The code can be found at https://github.com/AI4Science-Westlake U/CL_Diff Phy Con. |
| Open Datasets | No | We follow instructions in Wei et al. (2024) to generate a 1D Burgers equation dataset. Specifically, we use the finite difference method (FDM) to generate trajectories in a domain of space range x [0, 1] and time range ̓[0, 1], with random initial states and control sequences following certain distributions. The space is discretized into 128 cells and time into 10000 steps. We generated 90000 trajectories for the training set and 50 for the testing set. |
| Dataset Splits | Yes | We generated 90000 trajectories for the training set and 50 for the testing set. We generated 40,000 training trajectories for the large domain control setting, and 30,000 for the boundary control setting. During inference, for both settings, besides a fixed map (FM) evaluation mode where test samples use the same obstacles configuration with training trajectories, we also introduce a random map (RM) mode where the obstacles configuration varies. In this mode, all 50 test samples use the same obstacle configuration with training trajectories. |
| Hardware Specification | Yes | Table 1: Comparison results on 1D Burgers equation control. The average control objective ( J ) and inference time (averaged across all settings) are reported, using a single NVIDIA A100 80GB GPU with 16 CPU cores. Table 2: Comparison results on 2D incompressible fluid control. The average control objective ( J ) and inference time (averaged across all settings) are reported, using a single NVIDIA A6000 48GB GPU with 16 CPU cores. |
| Software Dependencies | No | The paper mentions "Phiflow solver Holl et al. (2020)" for fluid dynamics simulation, and refers to "U-Net (Ronneberger et al., 2015)" and "3D U-net Ho et al. (2022)" as architectures. However, it does not provide specific version numbers for any software libraries, frameworks, or languages used for implementation beyond these mentions. |
| Experiment Setup | Yes | Both models have T = 900 diffusion steps. The DDIM (Song et al., 2020) sampling we use only has 30 diffusion steps with the hyperparameter ̣= 1. Hyperparameters of models and training are listed in Table 3. Both models have T = 600 diffusion steps. The DDIM (Song et al., 2020) sampling we use has 75 diffusion steps with the hyperparameter ̣= 0.3 for the large domain control setting and 120 diffusion steps with the hyperparameter ̣= 0.3 for the boundary control setting. Hyperparameters of models and training are listed in Table 5. |