{"id":1137041,"date":"2025-04-18T00:00:54","date_gmt":"2025-04-18T07:00:54","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&p=1137041"},"modified":"2025-04-18T00:00:56","modified_gmt":"2025-04-18T07:00:56","slug":"peace-project-unlocks-ai-applications-in-geology-using-geomap","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/peace-project-unlocks-ai-applications-in-geology-using-geomap\/","title":{"rendered":"PEACE project unlocks AI applications in geology using GeoMap"},"content":{"rendered":"\n

When earthquakes hit, they often come with little warning. Each year, about 500,000 earthquakes ripple through the Earth\u2014some are felt, many aren\u2019t, but all are part of a complex and dynamic system that geologists are trying to understand. Earthquakes are notoriously difficult to predict, but a new interdisciplinary effort is using AI to bring us closer. At Microsoft Research Asia, researchers have developed an AI-agent system designed to help geologists understand geologic maps\u2014tools foundational to long-term earthquake assessment and risk analysis. Known as PEACE<\/a> (emp<\/strong>owering ge<\/strong>ologic ma<\/strong>p holistic<\/strong> unde<\/strong>rstanding), the initiative introduces two tools: GeoMap-Bench (opens in new tab)<\/span><\/a>, a benchmark dataset for evaluating AI performance on geologic maps, and GeoMap-Agent (opens in new tab)<\/span><\/a>, an AI-powered agent to read and analyze them\u2014work recently accepted at CVPR 2025.<\/p>\n\n\n\n

Together, they represent a major step forward in bringing vision language models (VLMs) into the realm of earth sciences\u2014and a promising way to speed up and scale geological insights that once took teams of experts days or even weeks to compile.<\/p>\n\n\n\n

Why geologic maps matter<\/h2>\n\n\n\n

Geologic maps are essential tools in earthquake prediction, mineral exploration, infrastructure planning, and environmental assessment. They depict layers of Earth\u2019s crust\u2014faults, rock formations, and tectonic boundaries\u2014often in intricate visual detail. When cross-referenced with tectonic activity and data on the crust\u2019s stability, these maps help scientists assess seismic risk and understand long-term geologic trends.<\/p>\n\n\n\n

But interpreting these maps isn\u2019t simple. They\u2019re dense with symbols, annotations, and embedded knowledge that even experienced geologists must study carefully. And while AI has made strides in image analysis and text comprehension, it still struggles to handle the complex, multimodal nature of geologic data.<\/p>\n\n\n\n

In many ways, geologic maps are languages in themselves. They\u2019re rich in meaning, but both technical and domain-specific understanding are needed to read them well.<\/p>\n\n\n\n

GeoMap-Bench: Establishing a benchmark for geologic map understanding<\/h2>\n\n\n\n

AI applications in geology are still in their early stages. To close this gap, Microsoft researchers, in collaboration with the Chinese Academy of Geological Sciences and Wuhan University, created GeoMap-Bench\u2014the first benchmark tailored for evaluating geologic maps. The team began by identifying five key capabilities for effective map interpretation:<\/p>\n\n\n\n

    \n
  1. Extracting:<\/strong> Can the model retrieve basic information, like title, scale, and, longitude and latitude coordinates?<\/li>\n\n\n\n
  2. Grounding:<\/strong> Can it find specific components when prompted by names or intensions?<\/li>\n\n\n\n
  3. Referring:<\/strong> Can it match names to their corresponding properties, such as identifying the rock name by its legend color?<\/li>\n\n\n\n
  4. Reasoning:<\/strong> Can it perform high-level logical tasks that require connecting information across components or incorporating external knowledge?<\/li>\n\n\n\n
  5. Analyzing:<\/strong> Can it comprehensively interpret a given topic on the map and provide detailed and meaningful insights from various perspectives?<\/li>\n<\/ol>\n\n\n\n

    Based on these competencies, the team created 25 evaluation tasks (Figure 1) and compiled a dataset of over 100 representative geologic maps\u2014sourced from over 7,000 maps provided by the China Geological Survey and the United States Geological Survey. From this, they generated more than 3,000 standardized questions to rigorously test multimodal AI models.<\/p>\n\n\n\n

    \"chart,
    Figure 1. Distribution of question types in GeoMap-Bench for 25 evaluation tasks<\/figcaption><\/figure>\n\n\n\n

    \u201cIf AI models can achieve more accurate geologic map interpretation, they will have a profound impact on surveying, geographic information mapping, cartography, navigation and positioning services, and even autonomous driving,\u201d said Zhipeng Gui, director at Wuhan University\u2019s School of Remote Sensing and Information Engineering. “It could signal a shift in geographic exploration and innovation.\u201d<\/p>\n\n\n\n

    GeoMap-Agent: Accelerating AI-driven geologic map analysis<\/h2>\n\n\n\n

    Building GeoMap-Bench was the first step. Researchers then used the benchmark to evaluate leading vision language models (VLMs), revealing key limitations in how current AI models interpret geologic maps. Their analysis identified three major challenges:<\/p>\n\n\n\n