Large Language Model-assisted Surrogate Modelling for Engineering Optimization

Engineers often utilize data-driven surrogate models in optimization to partially replace the costly computational simulations of physics-based models. However, different aspects affect the accuracy and prediction capability of the surrogate models, e.g., the dimensionality of the data and nonlinearities in the mapping between input and output. Learning well-performing models by selecting appropriate techniques to fit the data benefits from the machine learning and data science expertise of the engineer which may vary depending on the application domain. Recently, large language models (LLMs) have shown promising capabilities to support humans through natural language-based interfaces in approaching technical problems, as well as to co-develop software and to democratize domain knowledge. In this paper, we utilize ChatGPT 4 to co-develop a framework to select and train surrogate models for engineering optimization tasks. More specifically, we interact with ChatGPT to outline a process and software to support the selection and application of regression techniques based on characteristics of the available data and target application. We evaluate the developed methodology on synthetic and realistic engineering optimization data and problems. In our experiments, we demonstrate that the models obtained through the methodology developed with ChatGPT achieve comparable performance in regression and optimization tasks than observed in existing works in the literature. Hence, despite some limitations, such as missing updates of available software libraries, LLMs can support less experienced engineers to solve surrogate-assisted optimization problems more efficiently by providing insights on the application data and software for deploying surrogate models.