Optimizing the rapid assessment of microclimate environments in settlement heritage spaces: a large language model-enabled framework

Large language models (LLMs) offer new potential for microclimate analysis in traditional settlements—but their predictive accuracy remains unclear, especially for indoor environments and complex settlement spaces. To test this, we compared ChatGPT-4’s predictions with ENVI-met simulation results across diverse outdoor and indoor settings in traditional settlements. Key findings: (1) LLMs predict outdoor microclimate regulation accurately but perform poorly indoors; (2) they achieve higher accuracy in simple open spaces (e.g., farmland, squares); (3) Prediction accuracy improves significantly under extreme heat—especially in summer. The research attempts and empirical studies on the intervention of LLMs in microclimate prediction not only broaden the application potential of LLMs in the fields of architecture and environmental engineering, but also lay the foundation for the integration of LLMs into future microclimate research.