PINP: Physics-Informed Neural Predictor with latent estimation of fluid flows
Authors: Huaguan Chen, Yang Liu, Hao Sun
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To evaluate the effectiveness of the proposed method, we conducted a comparative study against several state-of-the-art models for predicting future observed data. The evaluation was performed using both numerically simulated 2D and 3D datasets, as well as real-world nowcasting benchmarks. The predictive accuracy of the models was quantitatively assessed across various scenarios. Experimental results demonstrate that our approach achieves SOAT performance in predicting observed data, while simultaneously estimating multiple latent physical quantities for interpretability, and exhibiting better temporal extrapolation and spatial generalization capabilities (Figure 1). |
| Researcher Affiliation | Academia | Huaguan Chen1, Yang Liu2, Hao Sun1, 1Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China 2School of Engineering Science, University of Chinese Academy of Sciences, Beijing, China Emails: EMAIL; EMAIL; EMAIL |
| Pseudocode | No | The paper describes methods using mathematical equations and textual explanations, but it does not contain any clearly labeled pseudocode blocks or algorithms. |
| Open Source Code | Yes | Our source codes are available in the following Git Hub repository: https://github.com/intell-sci-comput/PINP. |
| Open Datasets | Yes | SEVIR. The Storm EVent Image Ry (SEVIR) dataset (Veillette et al., 2020) contains meteorological events across the United States between 2017 and 2019. |
| Dataset Splits | Yes | For nowcasting, we selected the NEXRAD radar VIL composites. The VIL data has a spatial resolution of 1 km and is recorded at 5-minute intervals. Following the approach of Gao et al. (2022b), we use 65 minutes of VIL data (13 frames) to predict up to 60 minutes ahead (12 frames) for precipitation nowcasting. Due to computational limitations, we downsample the spatial resolution to 96 96. This downsampled dataset includes 35,718 training samples, 9,060 validation samples, and 12,159 test samples. |
| Hardware Specification | No | The paper mentions "GPU memory consumption" in Figure 10, but does not provide specific hardware details such as GPU models, CPU models, or other computer specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions using COMSOL software and ΦFlow for data generation, but does not provide specific version numbers for these or any other software libraries/frameworks (e.g., Python, PyTorch, CUDA) used for implementing and running their model. |
| Experiment Setup | Yes | Table S3: Hyperparameter and Training setup Type Name Meaning Value Hyperparameter Mask1 First-order gradient operator mask Eq. S3 Hyperparameter Mask2 Second-order gradient operator mask Eq. S4 Training setup LR Learning Rate 1e-3 Training setup Epoch Number of training epochs 100 Training setup Optimizer Type of model optimizer Adam Training setup Scheduler Schedule the learning rate of the optimizer Step LR Training setup Batch Size the number of samples processed together (2, 2, 2, 32) () : Value is different for each dataset, in the order of 2D fluid, 2D smoke, 3D smoke, and SEVIR. |