TurtLES benchmark: every 3D turbulence surrogate fails uniquely

Every machine-learning surrogate tested against a new 14 TB benchmark of three-dimensional turbulent flow ran faster than the reference solver, and every one broke in a different way. That is the finding of TurtLES, a dataset and benchmark paper accepted to the NeurIPS 2026 Datasets and Benchmarks track from authors including Paola Cinnella and Patrick Gallinari.

The dataset holds 400 trajectories of 400 time steps each, with 3 to 11 million points per frame, computed via implicit large-eddy simulations of cylinder wakes in the shear-layer transition regime. Across four baselines, "every surrogate provides a net speedup over the solver, ranging from ∼3× for GINO to ∼16× for UPT," the paper reports in GPU-seconds.

The failure modes were not the same. UPT "begins with a moderate error but drifts rapidly over time, reaching a VRMSE of 5.12." Transolver++ "strongly overpredicts spectral energy, especially at intermediate and high wavenumbers." GINO "did not converge within the shared training budget." The abstract groups these under temporal instability, loss of small-scale energy, and challenges in learning on large irregular domains.

Prior benchmarks had captured at most one of realistic 3D turbulence, long rollouts, and industrial-scale irregular meshes. TurtLES combines all three, which is the gap the authors say they set out to close.