PEACE

PEACE (emPowering gEologic mAp holistiC undErstanding) is a framework that injects geologic-domain expertise into multimodal large language models through three modules: Hierarchical Information Extraction (HIE) for managing high-resolution maps with many components, Domain Knowledge Injection (DKI) for geological reasoning, and Prompt-enhanced Question Answering (PEQA) tuned for spatial reasoning. The release includes GeoMap-Bench, the first benchmark for evaluating MLLMs on geologic-map understanding across extraction, referring, grounding, reasoning, and analysis. GeoMap-Agent scores 0.811 on the benchmark compared to 0.369 for GPT-4o.

PEACE addresses a real gap in earth-science AI: general-purpose MLLMs lack the cartographic and geological vocabulary required to interpret geological maps, which are dense, conventional, and information-rich. By formalizing the task and shipping a benchmark, the work also gives the broader community a yardstick for progress. The downstream applications — disaster detection, mineral and resource exploration, civil engineering — touch some of the highest-stakes decisions in earth science and exemplify Microsoft AI for Science’s pattern of building AI tools shaped to the data modalities and expert knowledge of each natural-science discipline.