GeoEvolve: Automating Geospatial Model Discovery via Multi-Agent Large Language Models

Peng Luo, Xiayin Lou, Yu Zheng, Zhuo Zheng, Stefano Ermon

Keywords: Large Language Models, AI for Science, Geospatial modeling

Abstract:

efficient AI-for-Science discovery.
Geospatial modeling provides critical solutions for pressing global challenges such as sustainability and climate change. Existing large language model (LLM)–based algorithm discovery frameworks, such as AlphaEvolve, excel at generic code evolution but lack the domain knowledge required for complex geospatial problems. We introduce GeoEvolve, a multi-agent LLM framework that couples evolutionary search with dynamic geospatial domain knowledge. GeoEvolve operates in nested loops: an inner code evolver generates candidate solutions, while an outer agentic controller—supported by Automated Knowledge Construction and Code-to-Formula agents—queries a Dynamic GeoKnowRAG module to inject theoretical priors. This architecture addresses the challenges of spatial heterogeneity and temporal non-stationarity. We evaluate GeoEvolve on three classical tasks: spatial interpolation (Kriging), uncertainty quantification (GeoCP), and spatial regression (GWR). Across 9 datasets, GeoEvolve discovers novel algorithms that incorporate geospatial theory. It achieves significant gains, such as a 29.5\% increase in regression $R^2$ and a 13–21\% reduction in interpolation error. Furthermore, extensive ablation studies confirm GeoEvolve’s robustness across diverse foundation models (GPT, Gemini, Qwen) and its spatiotemporal generalizability, validating that domain-guided retrieval is essential for stable evolution.
Collectively, these results offer a scalable path toward trustworthy, automated geospatial modeling, opening new avenues for efficient AI-for-Science discovery.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 19504