Model-based system engineering is a promising approach to achieve the design goal of green aviation by running virtual simulation to reduce energy consumption and emission. The virtual simulation of airborne systems encompasses the logic simulation of avionics and electromechanical controllers, as well as functional simulations of subsystems such as power system and hydraulics system. Certain complex electromagnetic and fluid equations involved in this simulation possess characteristics of high computational costs and small convergence solving steps. Such inconsistency in the timescales of the entire simulation system results in a delay in the overall numerical simulation time. A common solution to address this issue is to employ surrogate models for model reduction. The current model reduction methods used in airborne system integrated simulations in practical engineering tasks remain at the interpolation table stage, with low data complexity. However, most airborne system or module models are multidimensional, making it challenging to construct high-dimensional surrogate models. To tackle these challenges, this paper proposes a semi-automatic model reduction method based on large language models and existing mature fitting algorithms. A complete toolchain for model reduction has been developed, firstly, utilizing large language models to assist in selecting fitting algorithms and constructing fitted models. Then, training and deployment of models are conducted using an artificial intelligence model development platform, which is integrated into the GvSimLab platform to connect to other airborne system models for joint simulation. Taking a particular IPMSM module(Internal Permanent Magnet Synchronous Motor) in airborne VFG system(Variable Frequency Generator system) as an example, compared to the currently used interpolation tools, this toolchain improves the accuracy of simulating non-sample data by 15%. Compared with the Simulink model, this toolchain can elevate the model-in-loop solving time from 30s per thousand steps to nearly real-time at 0.8s with an accuracy of 96.4%.