AI recovers the disturbances that a fluid flow amplifies

A PNAS study published on 16 October turns a learned simulator into a tool for probing physical disturbances. Researchers train a neural network on state-to-state evolution, then differentiate it to estimate how small changes propagate. The network supplies sensitivity information normally obtained from the governing equations.[1][2]

A stable flow can amplify a disturbance

Fluid stability involves different timescales. A disturbance can grow before eventually dying away. In 1993, Trefethen and colleagues showed why inspecting decay rates alone can miss that behaviour: separately decaying flow patterns can combine to produce temporary growth.[3]

Picture a small vortex carrying faster fluid towards a slower region, and slower fluid the other way. The exchange can build elongated streaks with different speeds. A related question is how strongly a flow responds when an external disturbance keeps oscillating. That depends on the input’s spatial pattern and frequency. A flow can therefore be stable in the long run yet highly receptive to particular forcing.[3]

For three-dimensional channel flow at Reynolds number 2,000, the new study used 60 simulated trajectories, each containing 400 snapshots, compressed into 100 spatial patterns. The unperturbed flow was linearly stable.[1]

Preprocessing produced an unphysical response peak at zero frequency, reduced by filtering modes or training on velocity perpendicular to the channel walls. Rapidly decaying patterns were missed. The tests assumed observations of the full system state.[1]

The learned system nevertheless recovered the physically relevant response peak at dimensionless angular frequency 0.38, identifying a sensitive frequency and the flow pattern it excites.[1]

Report an error or suggest a correction