Limitation or failure · Physics · Mathematics · Computer science
Systematic review finds weak baselines in most claimed ML fluid-solver speedups
A systematic review identifies weak numerical comparisons in 60 of 76 fluid-related ML-solver papers claiming to beat standard methods.
Summary
McGreivy and Hakim reviewed fluid-related PDE acceleration claims and found 60/76 used weak baselines. They identify two main problems: comparing methods at unequal accuracy and comparing against inefficient numerical methods. They also document reporting biases that favor positive results. The journal publication was September 25, 2024, following a July 9 preprint.
AI role
The reviewed ML systems approximated forward fluid-related PDE solutions and claimed runtime or efficiency improvements over numerical solvers.
Narrative role
Fills a historical physics gap with evidence that reported simulation speedups can shrink or disappear under fair numerical comparisons, directly qualifying the process-acceleration story.
Caveat
The review covers selected fluid-related forward problems and excludes PINNs from its main systematic sample. Baseline classifications involve expert judgment; the result is not evidence that every ML solver lacks value or that inverse problems cannot benefit.