Abstract Subduction zones are central to Earth's tectonic evolution. Previously, identification and tracking of subduction zones in numerical mantle convection models have relied on methods such as detecting gradients in horizontal surface velocity. However, these approaches require the use of arbitrary thresholds and often lack sufficient spatial context, making robust and consistent detection challenging. This study adapts an established deep‐learning workflow for subduction zone (SZ) detection using Fully Convolutional Networks (FCNs), which perform semantic segmentation on RGB images constructed from temperature, vertical velocity, and fineness fields derived from 2D geodynamic simulations. Trained on a curated data set with high‐resolution ground truth masks, the FCN effectively identifies SZs with improved spatial coherence and robustness compared to traditional surface velocity convergence methods. Incorporating an attention‐based architecture and a custom loss function, the FCN captures complex SZ morphologies, reduces false detections, and enables consistent tracking of SZ evolution over time. Statistical analysis of tracked SZs reveals that those forming near continental margins are longer‐lived and less sensitive to mantle convective vigor, while oceanic SZs are shorter‐lived and strongly influenced by convective vigor. These findings support the hypothesis that continental lithosphere plays a stabilizing role in early Earth subduction dynamics. This FCN‐based method provides a threshold‐free, data‐driven alternative for SZ detection and offers a powerful tool for analyzing subduction processes in Earth and other planetary interiors.
Choi et al. (Sun,) studied this question.