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..
Hierarchical Implicit Neural Emulators
Authors: Ruoxi Jiang, Xiao Zhang, Karan Jakhar, Peter Y. Lu, Pedram Hassanzadeh, Michael Maire, Rebecca Willett
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on turbulent fluid dynamics show that our method achieves high short-term accuracy and produces long-term stable forecasts, significantly outperforming non-hierarchical autoregressive baselines while adding minimal computational overhead. |
| Researcher Affiliation | Academia | 1Department of Computer Science, The University of Chicago 2Department of Statistics, The University of Chicago 3Department of Geophysical Sciences, The University of Chicago 4Department of Physics, The University of Chicago 5Artificial Intelligence Innovation and Incubation Institute, Fudan University 6Department of Mechanical Engineering, Rice University 7Department of Electrical and Computer Engineering, Tufts University |
| Pseudocode | No | The paper describes the methodology using prose and mathematical formulations in Sections 3 and 4, but does not include a dedicated pseudocode or algorithm block. |
| Open Source Code | Yes | The codebase is available at this link1. 1https://github.com/roxie62/Hierarchical-Implicit-Neural-Emulators |
| Open Datasets | Yes | Training data is generated using py2d [52] on a 512 × 512 grid, with 18,000 ω snapshots. For analysis in this paper, the data is downsampled to a 256 × 256 grid, except in Section 5.3, which discusses ablation studies on different datasets and resolutions. [52] Karan Jakhar, Rambod Mojgani, Moein Darman, Yifei Guan, and Pedram Hassanzadeh. py2d: High-performance 2D Navier-Stokes solver in python (version 0.1), 2024. URL https://github.com/envfluids/py2d. |
| Dataset Splits | No | Given N + 1 consecutive observations {u0, u1, . . . , u N}, our goal is to learn a neural emulator fθ that approximates the underlying dynamics and predicts future states. No explicit train/test/validation split of the 18,000 snapshots is mentioned for the experiments. |
| Hardware Specification | Yes | For all experiments, we use the Adam W optimizer with learning rate at 3 × 10^-4. We conduct experiments on NVIDIA A100, H100, and L40S GPUs. |
| Software Dependencies | No | We process the raw data generated from py2d [52] solver. Our emulator predicts vorticity at 0.05 time intervals... For all experiments, we use the Adam W optimizer with learning rate at 3 × 10^-4. No specific version numbers for software libraries or frameworks (e.g., PyTorch, TensorFlow) are provided. |
| Experiment Setup | Yes | Our emulator predicts vorticity at 0.05 time intervals... We normalize the raw simulation data to approximate a standard normal distribution. Specifically, for the dataset with Reynolds number 10000, we apply a dividing factor of 10, while for the dataset with Reynolds number 5000, we use a dividing factor of 6. For all experiments, we use the Adam W optimizer with learning rate at 3 × 10^-4. |