Parameter optimization of nuclear thermal-hydraulic components involves coupled variables and complex engineering constraints, making it challenging to identify high-performance feasible designs within a limited evaluation budget. This study proposes a large language model-driven parameter optimization method (LLM-DO) that organizes design variables, physical constraints, and a deterministic evaluation interface into a structured interaction protocol, enabling closed-loop search through candidate generation, evaluation feedback, and design adjustment. A printed circuit heat exchanger (PCHE) recuperator in the S4 space nuclear reactor is considered. Volumetric power density is maximized using five design variables: the cold-side and hot-side channel diameters, transverse and longitudinal pitch-to-diameter ratios, and total number of heat transfer units. Constraints are imposed on the hot- and cold-side pressure drops, the minimum wall thickness, and the lower bounds of the inlet Reynolds numbers. A one-dimensional subchannel reduced-order model serves as the deterministic evaluator. LLM-DO is compared with an improved genetic algorithm (IGA) and Gaussian process Bayesian optimization (GP-BO) under the same evaluation model, design space, constraint set, and objective function, using their respective evaluation budgets and stopping criteria. Ten independent runs are performed for each method. Under the complete method-specific budgets, LLM-DO achieves a maximum volumetric power density of 1,380,080 W/m³ and a ten-run mean of 1,125,226 W/m³, 5.0% above the GP-BO mean; this difference is not statistically significant. A supplementary comparison under a common 50-FE upper cap yields mean values of 313,583, 817,612, and 1,125,226 W/m³ for IGA, GP-BO, and LLM-DO, respectively. LLM-DO uses 10–47 FE per run because of early stopping, so the cap does not equalize completed evaluations. Its recorded parameter-adjustment rationales provide a potentially useful, reviewable engineering audit trail, but are not evidence of internal model reasoning or physical causality. Comparison of the one-dimensional predictions for the representative best design with three-dimensional computational fluid dynamics (CFD) results shows a heat-transfer-rate deviation of no more than 4% and cold-side and hot-side pressure-drop deviations of 0.4% and 14.8%, respectively. The CFD hot-side pressure drop is 96.55 kPa and remains within the prescribed constraint. The results indicate potential for constrained PCHE parameter optimization under the present model and settings, while cross-model and cross-condition applicability requires further validation.

large language model,parameter optimization,printed circuit heat exchanger,space nuclear reactor,constrained optimization