Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Breaking the Discretization Barrier of Continuous Physics Simulation Learning
Authors: Fan Xu, Hao Wu, Nan Wang, Lilan Peng, Kun Wang, Wei Gong, Xibin Zhao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments demonstrate that Co PS advances the state-of-the-art methods in space-time continuous modeling across various scenarios. The source code is available at https://github.com/Sunxkissed/Co PS. Section 4: Experiment. |
| Researcher Affiliation | Academia | 1University of Science and Technology of China 2Tsinghua University 3Beijing Jiaotong University 4Southwest Jiaotong University 5Nanyang Technological University |
| Pseudocode | Yes | B Pseudocode of Co PS. Algorithm 1 Algorithm workflow of Co PS. |
| Open Source Code | Yes | The source code is available at https://github.com/Sunxkissed/Co PS. |
| Open Datasets | Yes | To comprehensively illustrate the property and efficacy of the extrapolations obtained from Co PS, we conduct experiments on diverse synthetic and real-world datasets. For synthetic datasets, we first choose Navier-Stokes [29] and Rayleigh Bรฉnard Convection [43], which are directly generated by numeric PDE solvers. Then, we select Prometheus [48], which is a large-scale combustion dataset simulated with industrial software. For real-world datasets, we choose Weather Bench [36], a dataset for weather forecasting and climate modeling. We also select Kuroshio [49], which provides vector data of sea surface stream velocity from the Copernicus Marine Service. |
| Dataset Splits | Yes | For synthetic data like Navier-Stokes, the train and test sets differ only by their initial conditions. The samples are partitioned in a 7:2:1 ratio into training, validation, and test sets. For real-world data like Weather Bench, we use the historical ERA5 global atmospheric reanalysis data. The data was partitioned strictly by date to prevent any data leakage from the future into the training process. Specifically, the train set uses data from 1979-2018, the validation set from 2019, and the test set from 2020-2022. ... Kuroshio: This dataset covers the period from 1993 to 2024, and we use data from 1993 2020 for training, while data from 2021 2024 for validation and testing. |
| Hardware Specification | Yes | To ensure fairness, we conducted all experiments on an NVIDIA-A100 GPU using the MSE loss over 200 epochs. |
| Software Dependencies | No | To ensure fairness, we conducted all experiments on an NVIDIA-A100 GPU using the MSE loss over 200 epochs. We used Adam optimizer with a learning rate of 10 3 for training. The batch size was set to 16. |
| Experiment Setup | Yes | To ensure fairness, we conducted all experiments on an NVIDIA-A100 GPU using the MSE loss over 200 epochs. We used Adam optimizer with a learning rate of 10 3 for training. The batch size was set to 16. The MFN is configured with 5 layers and uses Re LU as the activation function to ensure efficient feature extraction. The hidden dimension is set to 128. Then the multiscale graph ODE module utilizes a Runge-Kutta ode solver for numerical integration, with a time step size of 0.25 to ensure accurate modeling of temporal continuity. Further, the neural auto-regressive correction module performs corrections per integer time step. For this module, the Conv2d layer is downsampled to half the resolution, while the Up Conv2d layer restores the grid to the original resolution. The parallel Group Conv2d operations are implemented with filter sizes of 3 3, 5 5, and 7 7. For inference, the correction weight ฮป is set to 0.5 to balance correction strength and model stability. Finally, in the decoder, we use a single step Gabor filter transformation to initial the features of query coordinates, and perform a 2-layers message-passing update to obtain the corresponding predictions. |