DEEP LEARNING-BASED TRACKING OF SUBDUCTION ZONES IN MANTLE CONVECTION MODELS

Accurate modeling of subduction initiation is essential for understanding the processes that drive plate tectonics and the formation of new subduction zones. While there have been advancements in numerical models of subduction zones, there has been limited research on methods to track subduction zones throughout model runs. In this study, we propose a new approach for tracking subduction zones over time in numerical mantle convection models using deep learning. Traditional deep learning models typically output a single class for the entire input, rather than segmenting it into several parts, which is necessary for subduction zone detection. To address this, we employ a fully convolutional network (FCN) model.

Our method involves converting temperature, fineness (defined as the inverse of grain size, a key parameter in grain-damage rheology), and vertical velocity fields into RGB images to improve classification accuracy. A semantic segmentation-based FCN model is then constructed to determine subduction zones on a per-pixel basis by exploring the relationships among these RGB images. This approach enables a more accurate and detailed analysis of subduction zone dynamics without the need for arbitrary thresholds to define what constitutes a subduction zone. By using efficient image segmentation based on deep learning, we enhance our detection and analysis of subduction zones, potentially providing valuable insights into the mechanisms that trigger subduction and influence global tectonic processes. The ability to accurately track subduction zones over time opens new avenues for investigating the initiation, evolution, and termination of subduction processes, as well as their broader implications for mantle dynamics and plate tectonics. This new technique represents a significant advancement in automated subduction zone analysis, offering geoscientists a powerful tool for investigating Earth's dynamic processes.