Explainable AI Reveals Wing Turbulence Structures Classical Theories Miss
ricardovinuesa · x · 2026-09-08
Ricardo Vinuesa and collaborators posted an arXiv paper using explainable deep learning to characterize coherent structures in turbulent wing flow—key to reducing aircraft fuel consumption.
Method: train a deep neural network to predict short-term flow evolution, then use Shapley-value attribution to identify the highest-relevance flow regions—defining structures by predicted evolution rather than classical kinematic criteria.
Key findings:
- As flow decelerates toward the trailing edge, predictive importance shifts from near-wall low-speed structures to 3D pairs of high- and low-speed fluid regions;
- These pairs match no classical structure family yet progressively dominate dynamically relevant flow, standing spanwise side-by-side around a near-vertical momentum interface with the strongest velocity jump;
- Their geometry is invariant in viscous units while volume expands approaching separation.
The revealed predictive organization opens new routes for turbulent flow control.
More from Research
- Harvard's Institutional Data Initiative releases 22M images and 1.9B annotated paragraphs — leppert · 2026-09-08
- Random Laser's Physics Does Feature Learning, Hitting 77.7% Accuracy From Just 10 Images — bravo_abad · 2026-09-08
- DeepMind open-sources Science Skills: agent skills integrating 30+ scientific databases — _philschmid · 2026-09-08
- Eric Topol: landmark studies upend brain aging dogma, age 50 is the inflection — EricTopol · 2026-09-08
- HistReNERF localizes historical photographs using NeRFs, accepted at ECCV 2026 — ducha_aiki · 2026-09-08
- ECCV 2025 talk: Analyzing internal representations of multimodal LLMs — abursuc · 2026-09-08