TurtLES Debuts 14 TB Benchmark for Neural Surrogates of 3D Turbulent Flows
Summary:TurtLES introduces a 14 TB dataset for evaluating machine-learning surrogates of three-dimensional turbulent cylinder wakes. Generated with high-fidelity implicit large-eddy simulations, it contains 400 trajectories of 400 time steps, with 3–11 million mesh points per frame and variation in geometry, Reynolds number, and angle of attack. The benchmark combines vortex-driven energy cascades, dense irregular meshes, and long-horizon temporal dynamics. Its autoregressive evaluations expose three p
TurtLES Takes Neural-Operator Evaluation Into Genuine 3D Turbulence
TurtLES introduces a roughly 14 TB dataset for testing whether machine-learning surrogates can accelerate computational fluid dynamics. The simulations were generated with high-fidelity implicit large-eddy simulation and contain 400 trajectories of 400 time steps each. Every frame has between 3 million and 11 million mesh points, with variations in geometry, Reynolds number, and angle of attack.
The central contribution is not a new surrogate architecture. It is a more demanding evaluation setting for neural operators. Existing datasets often emphasize one-step accuracy on smooth or regular grids, providing limited evidence about long-horizon behavior. TurtLES concentrates three difficult properties in one benchmark: active energy transfer, dense irregular geometry, and extended temporal dynamics.
Preserving Turbulent Dynamics
The dataset focuses on three-dimensional cylinder wakes in the shear-layer transition regime. These flows contain shear instability, vortex shedding, and three-dimensional vortex stretching, producing an active energy cascade across spatial scales. Unlike datasets that retain only mean fields or time-averaged statistics, TurtLES provides complete trajectories.
For a surrogate, this changes the learning problem from isolated frame prediction to trajectory evolution. A small error can shift vortex-shedding phase, alter the location of coherent structures, and then compound during autoregressive rollout. The irregular mesh further complicates neighborhood construction, geometric representation, and generalization across differently shaped domains.
Autoregressive Evaluation Exposes Three Failure Modes
The benchmark evaluates neural operators through autoregressive rollouts: the model predicts the next state, then uses its own prediction as input for subsequent steps. Unlike teacher forcing, this setup resembles deployment, where the surrogate must sustain a complete spatiotemporal trajectory without receiving clean ground-truth states at every step.
The study identifies three recurring weaknesses:
- Temporal instability: A model may remain accurate over a short horizon but develop oscillations, amplitude drift, or structural distortions during longer rollouts.
- Loss of small-scale energy: Visually plausible outputs can still be overly smooth, suppressing high-frequency structures that carry important turbulent statistics and progressively degrading the energy cascade.
- Difficulty on large irregular domains: Changes in mesh size, geometry, and boundary conditions affect both local representation and global rollout stability.
This makes one-step relative error insufficient. A practical evaluation should also cover error accumulation, vortex-shedding phase, spectral-energy retention, peak velocity, pressure and drag coefficients, and out-of-distribution behavior across geometry and Reynolds number.
Engineering Costs and Boundaries
A 14 TB corpus imposes real infrastructure requirements. Training typically needs streaming data access, randomized temporal windows, and chunked loading. Feeding every mesh point into a global Transformer or high-order graph network can quickly become constrained by attention complexity, memory bandwidth, and data movement.
Two cost questions must remain separate. The first is training cost. The second is whether inference is genuinely cheaper than CFD. Even if a neural operator predicts faster than a conventional solver, the complete workflow still includes mesh preparation, feature construction, boundary handling, and post-processing. Long autoregressive horizons can also increase both accumulated error and compute cost.
Moreover, implicit LES itself remains dependent on spatial and temporal resolution. TurtLES defines a more demanding learning target; it does not imply that a surrogate can replace experiments or DNS without qualification. Industrial decisions involving loads, fatigue, noise, or sensitive flow states still require experimental validation and uncertainty estimates.
Designing a Benchmark That Does Not Become a Leaderboard Artifact
A useful public dataset needs more than input tensors and targets. A defensible protocol should include in-distribution interpolation, geometry extrapolation, Reynolds-number extrapolation, and long-horizon evaluation. Training and test trajectories must be isolated so that adjacent time frames cannot leak across splits.
Metrics should include both one-step and multi-step errors, reported as distributions across time windows rather than only as aggregate averages. Spectral errors expose small-scale energy suppression, while physically integrated quantities test whether predictions remain useful for engineering decisions. Training cost, inference latency, convergence behavior, and qualitative failure cases should be disclosed alongside accuracy.
Practical Guidance for Developers
Start with a lightweight baseline such as a Fourier Neural Operator or another graph/mesh neural operator, and restrict each sample to manageable spatial chunks and temporal windows. Optimize multistep stability before extending the rollout horizon; otherwise training loss, temporal drift, and spectral smoothing become entangled.
Reserve Reynolds numbers and geometries that are completely excluded from tuning, and report both physical-integral errors and spectral errors. For online control applications, also measure latency, memory use, restart behavior, and distribution-shift detection rather than relying on offline average accuracy alone.
TurtLES's main value is a more industrially relevant test bed for three-dimensional turbulent surrogates. It may expose weaknesses that clean, short-horizon benchmarks hide, but that is precisely what makes it useful: researchers can measure whether a neural operator preserves turbulence, not merely whether it produces accurate next frames.
— END —Source: Hugging Face Daily Papers (2026-11-06)
Tags: #Neural Operators #Computational Fluid Dynamics #Turbulence Simulation #Scientific Machine Learning #PDE Surrogates
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments