Global-scale AI-powered prediction of hydrogen seeps

Natural hydrogen (H 2 ) holds promising potential as a clean energy source, but its exploration remains challenging due to limited knowledge and a lack of quantitative tools. In this context, identifying active H 2 seepage areas is crucial for advancing exploration efforts. Here, we focus on sub-circular depressions (SCDs) that often mark high H 2 concentration in soils, thought to correspond to deeper fluxes seeping at the surface, making them promising targets for exploration. Coupling open-access Google Earth© images and in-field H 2 measurement data, an artificial intelligence model was trained to detect seepage zones. The model achieves an average precision of 95%, detects and maps seepage zones in new regions like Kazakhstan and South Africa, highlighting its potential for global application. Moreover, preliminary spatial analyses show that geological features control the distribution of H 2 -SCDs that can emit billions of tons of H 2 at the scale of a sedimentary basin. This study paves the way for a faster and more efficient methodology for selecting H 2 exploration targets. Plain Language Summary Natural hydrogen is a promising clean energy source, but it remains difficult to explore due to a lack of accessible tools. In this study, we used free satellite images (Google Earth©) and in-field hydrogen measurements to identify specific surface features - small sub-circular depressions (SCDs) - that often mark areas where hydrogen is seeping from underground. We trained an artificial intelligence model to detect these depressions, using a dataset of confirmed hydrogen-emitting SCDs collected across five countries. Thanks to this diversity in the training data, the model can be applied at a global scale, having learned to recognize a wide variety of structures associated with hydrogen seepage. To validate its effectiveness, the model was tested on two random regions - in Kazakhstan and South Africa - and successfully identified over a thousand new potential hydrogen-emitting depressions. With an average precision of 95%, this tool offers a fast and reliable way to map natural hydrogen seepage zones, helping guide future exploration efforts worldwide. • A deep learning model identifies thousands of circular depressions potentially related to natural hydrogen seepages with 95% precision • Spatial and morphometric analyses of mapped hydrogen seepages suggest geological controls governing hydrogen migration • Artificial intelligence enables new exploration strategies at a global scale