Global machine-learning detection of submarine calderas

Submarine calderas are among the least documented volcanic structures on Earth, yet recent impactful events highlight their potential for severe geohazard consequences globally, including tsunamis, seafloor damage, and atmospheric impacts. Logistical limits have historically hindered global detection. Here we apply a machine-learning-based caldera detection algorithm to global bathymetry, enabling systematic identification of previously undetected submarine calderas. We found 78 calderas spanning different water depths (up to 5,600 m), diameters (up to 20 km) and tectonic settings (divergent, convergent, intraplate). A subset of eight shallow-water calderas, primarily within volcanic arcs, are identified as high-priority targets due to their high hazard potential. Our dataset fills a critical observational gap and offers a reproducible and upgradeable framework for submarine volcano characterization, geohazard assessment, and deep-sea exploration. These findings underscore the need to integrate submarine calderas into future global hazard modelling and monitoring strategies.