FMAgent: Autonomous multi-agent system for fluid mechanics

The application of large language models (LLMs) in science and engineering has shown immense potential. However, existing automated AI systems for scientific research mainly generate ideas/proposals and lack a closed loop from proposal generation to executable experiments or computations for real-world problems. This paper introduces FMAgent, an end-to-end autonomous multi-agent system that covers the entire research process, including proposal generation and executable engineering code for real-world computational problems. We propose a novel two-stage training paradigm that combines supervised fine-tuning with a multi-agent debate-based reinforcement learning for the proposal generator. This protocol effectively mitigates the “safe-answer” bias of single-judge LLMs, enabling a local 32B-parameter model to autonomously generate highly innovative and physically feasible proposals for engineering problems. These proposals are subsequently translated into fully debugged implementations via a robust “concept-to-code” module in FMAgent system with strict, domain-specific quality-assurance guardrails. To empirically validate the FMAgent system’s capacity of automatically doing research in real world, we take incompressible Navier–Stokes equations as an engineering problem example. The system autonomously designed PANSEDGE, a novel neural operator with elliptical attention, multi-scale pyramid and state-space memory modules, which reflect physics-aware structural correspondences to rotational properties, multiscale features and temporal dynamics of fluid problems. Evaluated across one complex turbulent flow and four benchmarks including Pipe flow, Navier–Stokes (NS), compressible Airfoil problem and Darcy flow, PANSEDGE achieves competitive accuracy and generalization performance, rivaling or outperforming established human-designed architectures with high time efficiency. Ablation studies reveal that static system guardrails are indispensable for overcoming misalignment inherent to computational physics. This work demonstrates a paradigm shift toward automated research of AI agents themselves in engineering informatics with human-on-the-loop physical enhancement.