Problem

Traditional experimental fluid mechanics relies heavily on human expertise, limiting efficiency and thorough exploration of complex phenomena across high-dimensional parameter spaces. Current AI applications in fluid mechanics are mostly confined to numerical simulations, as physical experiments impose significant challenges like hardware integration, real-time feedback, safety risks, and combinatorial explosion of parameters. There's a need for an AI system that can reason under equipment constraints, interact with human researchers, and integrate automated control with intelligent parameter search to overcome these limitations and accelerate scientific discovery in experimental settings.

Method

The researchers propose an "AI Fluid Scientist" framework, a multi-agent system with virtual-real interaction, designed to automate the entire experimental fluid mechanics research cycle. The core of the system is a computer-controlled circulating water tunnel (CWT) capable of programmatic control over flow velocity, cylinder position, forcing parameters (frequency and amplitude), and data acquisition (displacement, force, torque). The framework operates in two modes: Human-in-the-Loop (HIL), where an LLM generates content and human experts provide judgment and feedback, and a fully autonomous end-to-end multi-agent system. The multi-agent system comprises six specialized LLM agents (hypothesis, experiment, hardware, analysis, evaluation, and manuscript agents) that coordinate to execute the workflow from hypothesis generation to publication-ready manuscript, with human involvement limited to initial hypothesis selection.

Results

The framework successfully reproduced literature benchmarks for Vortex-Induced Vibration (VIV) and Wake-Induced Vibration (WIV), with frequency lock-in within 4% and matching critical spacing trends. In HIL mode, the system discovered novel WIV amplitude response phenomena through iterative experimental campaigns, such as strong suppression windows at specific forcing frequencies and spacing-dependent critical Reynolds number transitions. The framework also autonomously explored different functional forms for fitting physical laws from data, ultimately finding that a neural network-based model achieved an R² of 0.7958, outperforming physics-based polynomial fittings by 31%. In the end-to-end autonomous mode, the multi-agent system executed five iterative experimental campaigns, totaling 222 configurations, to validate a hypothesis about nonlinear mode transitions in tandem cylinders, culminating in the autonomous generation of a publication-ready manuscript.

Implications

This work introduces a new paradigm for experimental fluid mechanics research by demonstrating the practical feasibility of AI-driven scientific discovery at scale. The AI Fluid Scientist significantly improves research efficiency, reduces human intervention in labor-intensive experiments, and enables systematic exploration of complex, multi-factor phenomena. The ability of the framework to autonomously discover new physical phenomena and derive accurate empirical laws using neural networks highlights the potential for AI to accelerate scientific advancement. However, the authors note a limitation in the evaluation agent's tendency towards overoptimistic validation assessments due to LLM hallucination, suggesting that the Human-in-the-Loop mode often yields more reliable results, emphasizing the need for specialized LLMs as discriminators.