Deep Neural Network-Based Inversion of Turbidites in Confined Basins

Turbidites generated by large earthquakes and other geological events are commonly preserved in small, topographically confined basins along active continental margins. Reconstructing flow conditions from these deposits is essential for assessing past hazards; however, existing inverse models have been validated only for unconfined settings, and their applicability to confined basins remains unclear. Here, we test a deep neural network (DNN)-based inverse framework using synthetic datasets generated in an idealized confined basin. Forward simulations reproduce key features of confined turbidity currents, including flow reflection and sediment ponding. Despite these complex dynamics, the trained DNN successfully reconstructs flow conditions with reasonable accuracy (SMAPE: 21.9–45.0%) using a limited number of training datasets. Notably, accurate inversion is achieved with as few as ten sampling points. These results demonstrate that inverse analysis of turbidites is feasible in confined basins even under sparse observational constraints. The proposed framework provides a practical basis for applying inverse modeling to natural deposits and offers a pathway toward linking turbidite records to the magnitude of their triggering events.