Deep learning 3D inversion of gravity data at volcanic seamounts: An example of Conception Bank volcano (Canary Islands)

Abstract Deep learning (DL)-based inversion significantly reduces computational time from hours to minutes compared with conventional methods. This efficiency motivates its application to satellite-derived gravity data to provide a robust preliminary model of subsurface density structures in inaccessible undersea volcanoes (seamounts) where no local gravity measurements exist. DL approaches enable capturing both lateral and vertical structural complexities that are challenging for conventional inversion methods. This study presents a 3D DL-based gravity inversion (3D fully convolutional neural network) applied to volcanic seamounts, using Conception Bank (Canary Islands) as a case study. The relative age of the Conception Bank seamounts and islands suggests it formed prior to the main Canary Islands volcanic edifices. However, it remains a geological enigma, despite its symmetric and extensive uplift, it never emerged as a subaerial volcanic island. First, the robustness of the proposed 3D DL inversion is validated using synthetic tests and a well-constrained volcanic setting at El Hierro Island, demonstrating its ability to accurately capture subsurface density variations. Application to Conception Bank reveals five distinct magma intrusions (one with lower certainty) extending to depths of approximately 8–9 km, with structural connections between some intrusions at 10–12 km depth along the Canary Island chain. The spatial distribution of these intrusions explains the observed symmetric bathymetric uplift and indicates a volcanic style characterized by moderate melt supply. Sensitivity tests indicate that the ML-based inversion is more robust to noise than conventional inversion methods. As typical for DL inversion, the fit to observed gravity data is moderate, and residuals can reach up to 20% of the input gravity signal when recovering complex density structures. Overall, DL-based inversion leverages learned mappings and architectural regularization to generate stable, physically plausible 3D density models, effectively mitigating the non-uniqueness and ill-posedness inherent in traditional gravity inversion.