AI fluid dynamics models exhibit hallucination in unstable flow

A recent study published on arXiv presents the inaugural systematic evidence of hallucinations in AI models related to fluid dynamics, focusing on the well-known issue of viscous fingering, which involves hydrodynamically unstable transport. The study highlights the difficulties AI faces in accurately modeling flow with instabilities due to the complex and rapidly changing fingering patterns. Researchers discovered that some solutions, while appearing visually plausible, are physically unrealistic, similar to hallucinations seen in large language models. These inaccuracies result in false fluid interfaces and reverse diffusion, breaching conservation laws. This issue stems from the spectral bias of AI models, which intensifies at elevated flow rates and viscosity differences. To address this, the team developed DeepFingers, a novel AI framework for fluid dynamics that integrates the Fourier Neural Operator to ensure balanced learning across various spatial modes.