From Zero to Turbulence: Generative Modeling for 3D Flow Simulation

Authors: Marten Lienen, David Lüdke, Jan Hansen-Palmus, Stephan Günnemann

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental For our experiments, we introduce a challenging 3D turbulence dataset of high-resolution flows and detailed vortex structures caused by various objects and derive two novel sample evaluation metrics for turbulent flows. On this dataset, we show that our generative model captures the distribution of turbulent flows caused by unseen objects and generates high-quality, realistic samples amenable for downstream applications without access to any initial state.
Researcher Affiliation Academia Marten Lienen, David Lüdke, Jan Hansen-Palmus, Stephan Günnemann Department of Informatics & Munich Data Science Institute Technical University of Munich, Germany {m.lienen,d.luedke,j.hansen-palmus,s.guennemann}@tum.de
Pseudocode No The paper describes the model (DDPM, U-Net, transformer) and training details in prose, but it does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Find code and data at https://cs.cit.tum.de/daml/generative-turbulence.
Open Datasets Yes To evaluate the viability of generative modeling for 3D turbulent flows, we have generated a challenging new dataset. The dataset consists of 45 simulations of an incompressible flow through a 0.4 m 0.1 m 0.1 m channel at 20 m s 1, which results in a Reynold s number of Re = 2 105, well into the turbulent regime. Find code and data at https://cs.cit.tum.de/daml/generative-turbulence.
Dataset Splits Yes We split the simulations into 27 for training, 9 for validation, and 9 for testing to verify that models can generalize to unseen objects in the flow.
Hardware Specification Yes All model times represent the minimum achieved time on an NVIDIA A100, and we measured the solver time with 16-core parallelism on an Intel Xeon E5-2630.
Software Dependencies No The paper mentions using Open FOAM as a CFD solver but does not specify its version. While PyTorch is referenced, no specific version of PyTorch or other software dependencies used for the experiments is provided in the main text or appendices.
Experiment Setup Yes We chose N = 500 and the log-linear signal-to-noise ratio (SNR) schedule from (Kingma et al., 2021) that scales βn such that the log-SNR of the data falls linearly from 1 103 to 1 10 5. We trained Turb Diff from 3 different seeds for 10 epochs with a batch size of 6. The optimizer is RAdam with a learning rate of 1 10 4 and and exponential learning rate decay to 1 10 6 over those epochs.