Large language model-assisted sensitivity analysis and parameter optimization in computational fluid dynamics

• Introduces a novel large language model-driven chain-of-thought (COT) framework that integrates computational fluid dynamics (CFD) simulations with sensitivity analysis and parameter optimization. • Proposes the OptMetaOpenFOAM system, enabling natural language inputs to automate complex CFD tasks and lowering the technical barriers for non-expert users. • Demonstrates the effectiveness of the approach on two benchmark cases: CounterFlowFlame2D and Hydrogen Combustion Chamber. • Employs advanced optimization techniques, including the L-BFGS-B algorithm for univariate optimization and the active subspace method for multivariable analysis. • Achieves significant simulation efficiency by converting concise natural language commands into extensive code for simulation, post-processing, and analysis. • Provides new insights into combustion dynamics and energy conversion processes, supporting advancements toward carbon-neutral energy systems. Integrating natural language interfaces with computational fluid dynamics (CFD) workflows presents transformative opportunities for both industry and research. In this study, we introduce OptMetaOpenFOAM, an innovative framework that employs a large language model (LLM)-driven multiagent architecture to automate sensitivity analyses and parameter optimization tasks in CFD via natural language instructions. By automating complex tasks via natural language inputs, the framework empowers nonexpert users to perform sensitivity analyses and parameter optimizations with markedly improved efficiency. The framework's efficacy is demonstrated through comprehensive testing across 11 distinct CFD tasks—including fluid flow, combustion, and heat transfer—originating from standard OpenFOAM tutorials and an external validation case involving hydrogen combustion chamber optimization. Remarkably, concise natural language commands (∼200 characters) successfully triggered elaborate computational sequences involving simulation setup, postprocessing, sensitivity analysis, and parameter optimization, translating into over 2,000 lines of automated code execution. These outcomes highlight the considerable potential of LLM-driven methodologies to advance CFD capabilities, offering enhanced accessibility, accuracy, and efficiency for both research-oriented and industrial CFD applications. The code is available at https://github.com/Terry-cyx/MetaOpenFOAM .