Measured acceleration · Physics · General science

Active learning compresses boundary-layer wind-tunnel exploration

Active-learning controlled physical wind-tunnel experiments that identified turbulence-profile equivalence regions with far fewer configurations

Summary

A Scientific Reports paper applied active learning to automated boundary-layer wind-tunnel experiments. The system used a computer-controlled roughness grid and active selection of physical configurations to learn turbulence-profile relationships that the authors estimate would require orders of magnitude more conventional experiments.

AI role

Active-learning algorithms selected informative wind-tunnel configurations in a closed loop with automated roughness-element actuation and measurement.

Narrative role

This is a historical-gap backfill for measured acceleration in physical experimentation, showing that active learning can reduce experiment counts outside chemistry and biology self-driving labs.

Caveat

The result is from a specialized wind-engineering apparatus and learning objective, so the reduction factors may not transfer directly to less automatable physical systems.

Event

Date: 2023-05-24
Timeline role: Supporting
Progress relevance: Process acceleration
Validation: Peer reviewed

Evidence

Peer-reviewed paper · Workflow metric

Observed metrics

phase 1 effort: 253 experiments in about 2 work weeks; The paper compares active-learning exploration against Monte Carlo or tensor-product space-filling designs for similar boundary resolution. Comparator: about 2500 experiments and 500 hours for a naive design.
phase 2 reduction: less than 100 experiments per case; The authors report about a 25-fold effort reduction in the second phase of the wind-tunnel study. Comparator: expected 2500 experiments.
high-dimensional search: 300 experiments plus about 200 follow-up experiments; The active-learning sensitivity analysis reduced a 10-dimensional random-field configuration before learning the equivalence region. Comparator: 10^10 configurations for a dense tensor-product design.

Source: Scientific Reports ↗