LLM-powered multi-agent system automates complex fluid dynamics simulations

Computational fluid dynamics (CFD) traditionally demands deep technical expertise and navigation of fragmented, intricate pipelines. Researchers have now developed Foam-Agent, a framework built on large language models that orchestrates the entire CFD workflow from a single plain-language instruction, making the process more accessible to non-experts.

The challenge#

CFD simulation—fundamental to engineering design, weather modeling, and physics research—remains locked behind steep learning curves. The process involves multiple stages: geometry meshing, configuration file generation, solver setup, and result analysis. Each stage depends critically on preceding steps, and subtle errors cascade throughout the pipeline. This complexity has historically excluded researchers and practitioners without specialized training.

How Foam-Agent works#

The system deploys six specialist agents, each handling a distinct phase: planning the overall workflow, meshing the geometry, synthesizing configuration files, executing simulations, reviewing results, and generating visualizations. Rather than treating the CFD knowledge as a single searchable corpus, Foam-Agent organizes domain information along four structural dimensions and retrieves relevant material based on the current workflow stage—improving the precision of what the LLM sees compared to conventional retrieval approaches.

A key innovation addresses a fundamental challenge: ensuring that generated configuration files remain consistent across the OpenFOAM case dependency graph. Instead of generating files independently, the system treats the relationships between files as a topology it must traverse. Each configuration is synthesized with full awareness of previously generated predecessors, maintaining logical coherence throughout.

When simulations fail, Foam-Agent doesn't start from scratch. Its trajectory-conditioned reviewer loop examines the accumulated history of errors and diagnoses from prior attempts, then applies targeted minimal edits to address only the reduced set of issues identified. This focuses each correction attempt rather than attempting wholesale reconfiguration.

Performance on benchmarks#

Testing on FoamBench—a suite of 110 basic-tier tasks—Foam-Agent achieved an 88.2% execution success rate without requiring expert intervention. On the more challenging out-of-distribution Advanced tier, it reached 62.5% success, demonstrating capability beyond its training distribution. The system also integrates with the Model Context Protocol, allowing external orchestrators to invoke its functionality.