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..
Generating Full-field Evolution of Physical Dynamics from Irregular Sparse Observations
Authors: Panqi Chen, Yifan Sun, Lei Cheng, YANG YANG, Weichang Li, Yang Liu, Weiqing Liu, Jiang Bian, Shikai Fang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate SDIFT on three physical systems spanning astronomical (supernova explosions, light-year scale), environmental (ocean sound speed fields, kilometer scale), and molecular (organic liquid, millimeter scale) domains, demonstrating significant improvements in both reconstruction accuracy and computational efficiency compared to state-of-the-art approaches. The code is available at https: //github.com/Ocean STARLab/SDIFT. ... Our contributions can be succinctly summarized as follows: ... 4) We demonstrate state-of-the-art performance across three distinct physical phenomena at various scales, highlighting the framework s generalizability and robustness. This work paves the way for more flexible, accurate and computationally efficient reconstructions of real-world physical systems, with broad implications for simulation and prediction in complex scientific and engineering tasks. ... 5 Experiment Datasets: We examined SDIFT on three real-world benchmark datasets, which span across astronomical, environmental and molecular scales. |
| Researcher Affiliation | Collaboration | Panqi Chen1 Yifan Sun1 Lei Cheng1, 2 Yang Yang3 Weichang Li1 Yang Liu4 Weiqing Liu4 Jiang Bian4 Shikai Fang1, 4 1 College of Information Science and Electronic Engineering, Zhejiang University 2 Zhejiang Provincial Key Laboratory of Multi-Modal Communication Networks and Intelligent Information Processing 3 College of Computer Science and Technology, Zhejiang University 4Microsoft Research Asia |
| Pseudocode | Yes | Algorithm 1 Message-passing Diffusion Posterior Sampling |
| Open Source Code | Yes | The code is available at https: //github.com/Ocean STARLab/SDIFT. |
| Open Datasets | Yes | Datasets: We examined SDIFT on three real-world benchmark datasets, which span across astronomical, environmental and molecular scales. (1) Supernova Explosion, temperature evolution of a supernova blast wave in a compressed dense cool monatomic ideal-gas cloud. We extracted 396 records in total, each containing 16 frames, where each frame has a shape of 64 64 64. We use 370 for training, randomly masking out 85% of the points in each record to simulate irregular sparse data. The remaining 26 records are reserved for testing.(https://polymathic-ai.org/ the_well/datasets/supernova_explosion_64/; (2) Ocean Sound Speed, sound speed field measurements in the pacific ocean. We extracted 1000 records of shape 24 5 38 76, using 950 for training (with 90% of points randomly masked) and reserving 50 for testing. (https: //ncss.hycom.org/thredds/ncss/grid/GLBy0.08/expt_93.0/ts3z/dataset.html). (3) Active Matter, the dynamics of rod-like active particles in a stokes fluid simulated via a continuum theory. We extracted 928 records, each of size 24 256 256, using 900 for training (with 90% of points randomly masked) and reserving 28 for testing.(https://polymathic-ai.org/ the_well/datasets/active_matter/) |
| Dataset Splits | Yes | (1) Supernova Explosion, ... We use 370 for training, randomly masking out 85% of the points in each record to simulate irregular sparse data. The remaining 26 records are reserved for testing. ... (2) Ocean Sound Speed, ... We extracted 1000 records of shape 24 5 38 76, using 950 for training (with 90% of points randomly masked) and reserving 50 for testing. ... (3) Active Matter, ... We extracted 928 records, each of size 24 256 256, using 900 for training (with 90% of points randomly masked) and reserving 28 for testing. |
| Hardware Specification | Yes | Sampling Speed: We compared the sampling speed of our method with Co NFi LD [5] on a NVIDIA RTX 4090 GPU with 24 GB memory; the results are shown in Tab. 2. |
| Software Dependencies | No | All the methods are implemented with Py Torch [42] and trained using Adam [43] optimizer with the learning rate tuned from {5e 4, 1e 3, 5e 3, 1e 2}. |
| Experiment Setup | Yes | All the methods are implemented with Py Torch [42] and trained using Adam [43] optimizer with the learning rate tuned from {5e 4, 1e 3, 5e 3, 1e 2}. For DEMOTE, we used two hidden layers for both the reaction process and entry value prediction, with the layer width chosen from {128, 256, 512}. For LRTFR, we used two hidden layers with layer width chosen from {128, 256, 512} to parameterize the latent function of each mode. We varied R from {16, 32} for all baselines. For Senseiver, we used 128 channels in both the encoder and decoder, a sequence size of 256 for the Qin array, and set the size of the linear layers in the encoder and decoder to 128. For Co NFi LD, we used a Conditional Neural Field module with a latent dimension of 256 for the Ocean Sound Speed and Active Matter datasets, and 1024 for the Supernova Explosion dataset. The diffusion model module was configured with 100 sampling steps. For our method, we first apply a functional Tucker model to decompose the tensor into factor functions and a core sequence. Each factor function is parameterized by a three-layer MLP, where each layer contains 1024 neurons and uses the sine activation function. The core sizes are set to 32 32 32, 3 12 12, and 48 48 for the Supernova Explosion, Ocean Sound Speed, and Active Matter datasets, respectively. These hyperparameters are carefully selected to achieve optimal performance. |