Source-linked AI summary
Single-snapshot machine learning for super-resolution of turbulence
Kai Fukami, Kunihiko Taira
TL;DR
Because machine-learning models for turbulence are commonly data-hungry, this paper asks whether one snapshot can suffice for extracting physical insights. It trains a super-resolution model on tiles from a single snapshot and reports reconstruction across Reynolds numbers for two-dimensional decaying turbulence and for turbulent channel flow.
Problem
The paper asks whether nonlinear machine learning can extract physical insights from turbulence using only one training snapshot rather than large datasets.
Source: §1 · p. 1Method
A DSC/MS super-resolution model is trained on local subdomains from one turbulent-flow snapshot and tested on independent snapshots.
Source: §2 · p. 2Results
Single-snapshot training reconstructs vortical structures across Reynolds numbers in two-dimensional decaying turbulence and reproduces fine-scale structures and statistics in channel flow.
Source: §1 · p. 2, §3.1 · p. 5, §3.2 · pp. 9, 11Takeaways & Limitations
Scale-invariant flow features and appropriate model architecture with physics embedding are important for extracting turbulent-flow insights without massive training datasets.
Source: §1 · p. 2The study uses an L2 norm for optimization; incorporating governing-equation knowledge into the cost function could better constrain the solution space.
Source: §2 · p. 3
Abstract
from arXiv · showhide
Modern machine-learning techniques are generally considered data-hungry. However, this may not be the case for turbulence as each of its snapshots can hold more information than a single data file in general machine-learning settings. This study asks the question of whether nonlinear machine-learning techniques can effectively extract physical insights even from as little as a {\it single} snapshot of turbulent flow. As an example, we consider machine-learning-based super-resolution analysis that reconstructs a high-resolution field from low-resolution data for two examples of two-dimensional isotropic turbulence and three-dimensional turbulent channel flow. First, we reveal that a carefully designed machine-learning model trained with flow tiles sampled from only a single snapshot can reconstruct vortical structures across a range of Reynolds numbers for two-dimensional decaying turbulence. Successful flow reconstruction indicates that nonlinear machine-learning techniques can leverage scale-invariance properties to learn turbulent flows. We also show that training data of turbulent flows can be cleverly collected from a single snapshot by considering characteristics of rotation and shear tensors. Second, we perform the single-snapshot super-resolution analysis for turbulent channel flow, showing that it is possible to extract physical insights from a single flow snapshot even with inhomogeneity. The present findings suggest that embedding prior knowledge in designing a model and collecting data is important for a range of data-driven analyses for turbulent flows. More broadly, this work hopes to stop machine-learning practitioners from being wasteful with turbulent flow data.
1 Introduction
Because machine-learning models for turbulence often require large training datasets, the paper asks whether one turbulent-flow snapshot can provide enough information for learning. It investigates single-snapshot super-resolution for two-dimensional decaying turbulence and three-dimensional channel flow.
1 Introduction
- Machine-learning models for turbulent-flow analysis commonly require far more training data than traditional analysis techniques.The cited applications include subgrid-scale modeling, reduced-order modeling, super-resolution, and flow control.Source: §1 · p. 1
- A turbulent snapshot may contain multiscale, scale-invariant structures that allow important flow features to be learned without large datasets.The study proposes subsampling and leveraging turbulent statistics, alongside architectures and learning formulations that incorporate physics.Source: §1 · p. 1
- Single-snapshot training can reconstruct flow across Reynolds numbers in two-dimensional decaying turbulence, and the study also examines three-dimensional channel flow.The authors conclude that large datasets are not necessarily needed for machine learning of turbulent flows.Source: §1 · p. 2
2 Approach
The approach trains a super-resolution model on subdomains from one turbulent-flow snapshot and tests it on independent snapshots. Its DSC/MS convolutional architecture is designed to accommodate turbulent length scales and vortical-structure invariances.
2 Approach
- The model learns to reconstruct a high-resolution flow field from low-resolution data, using subdomains from one snapshot and independent snapshots for testing.The model's weights are learned from the sampled subdomains.Source: §2 · p. 2
- The DSC/MS convolutional model combines downsampling and skip connections with multiple filter sizes to address invariance and a range of turbulent length scales.The DSC and MS subnetworks are internally connected through skip connections.Source: §2 · p. 2
- The study optimizes with an L2 norm, while incorporating governing-equation knowledge into the cost function is proposed as a way to better constrain solutions.Figure 1 identifies the interconnected DSC/MS model used for reconstruction.Source: §2 · p. 3
3 Results
With one-snapshot training, the model reconstructs two-dimensional vortical structures across Reynolds numbers and turbulent channel flow at a different Reynolds number. Results also show that scale-aware and moment-informed sampling can reduce the training tiles needed for reconstruction.
3.1 Example 1: two-dimensional decaying homogeneous isotropic turbulence; 3.2 Example 2: turbulent channel flow
- The model reconstructs two-dimensional vortical structures with 10–20% L2 error across Reynolds numbers and channel-flow fields with 5–8% error.The channel-flow model is tested on snapshots from an independent simulation.Source: Figure 3 · p. 4, §3.1 · p. 5, §3.2 · p. 9
3.1 Example 1: two-dimensional decaying homogeneous isotropic turbulence
- With increasing test Reynolds number, bicubic interpolation struggles across the vorticity distribution, whereas the model's distributions remain close to reference DNS.Bicubic interpolation deviates in the strong-rotation tail at Re0 = 80.4.Source: Figure 3 · p. 4, §3.1 · p. 5
- The model reproduces E(k) reasonably up to k ≈200 for Re0 = 80.4 and 177, and k ≈100 for Re0 = 442, while bicubic interpolation underestimates energy.At higher wavenumbers, weak correlation between low- and high-wavenumber components limits agreement.Source: §3.1 · p. 5, Figure 4 · p. 5
- Reconstruction improves as λtest/λsingle or the number of tiles increases; a smaller λsingle can achieve comparable accuracy with fewer tiles.A smaller λsingle generally corresponds to a high-Re training field with many vortical structures.Source: §3.1 · p. 6, Figure 6 · p. 7
- Third- and fourth-moment sampling with 250 tiles matches the error of random sampling with 2,000 tiles, while resolving finer vortices and shear layers.First- and second-moment sampling performs similarly to random tile selection.Source: §3.1 · p. 8, Figure 7 · p. 8
3.2 Example 2: turbulent channel flow
- In channel flow, the model matches velocity statistics across the wall-normal direction, while bicubic interpolation overestimates velocity in the viscous sublayer and part of the buffer layer.The model also reproduces the streamwise energy distribution and the decay of the spatial two-point correlation.Source: §3.2 · pp. 10–11, Figure 10 · p. 10, Figure 11 · p. 11
- Channel-flow reconstruction error increases with input noise; fine structures remain accurate up to γ ≈0.2, while large structures remain reconstructable with 50% noise.For channel flow, tiles are sampled from a single Reτ = 5200 snapshot; far from the wall, 500 tiles give reasonable reconstruction, while accurate near-wall reconstruction requires O(10^3) tiles.Source: §3.2 · pp. 9, 12, Figure 12 · p. 11, Figure 13 · p. 12
4 Concluding remarks
The study tests single-snapshot super-resolution on two-dimensional decaying isotropic turbulence and turbulent channel flow. It concludes that scale-invariant features and flow-informed model and data design can support learning from smaller turbulent-flow datasets.
4 Concluding remarks
- The DSC/MS convolutional super-resolution model learns low/high-resolution relations from local subdomains of one snapshot and is tested on independent simulations.Source: §4 · p. 12
- For two-dimensional decaying isotropic turbulence, statistical characteristics of a single snapshot support efficient preparation of training data.Source: §4 · p. 12
- For turbulent channel flow, single-snapshot training can learn low/high-resolution relations despite spatial inhomogeneity.Source: §4 · p. 12
- Scale-invariant features across Reynolds numbers, flow-appropriate physics-embedded architectures, and spatial and temporal redundancy inform smarter, smaller training sets.Source: §4 · p. 12