ShipNet: A Geometric Deep Learning Surrogate for Real-Time Ship Hydrodynamics

Researchers have developed ShipNet, a geometric deep-learning model that predicts ship hydrodynamics 550 times faster than traditional methods, achieving high accuracy in hull pressure and wave field predictions.

ShipNet, presented in a recent study, uses a regularized dynamic graph convolutional backbone and a multi-head decoder to predict hull-surface pressure distributions and far-field free-surface wave patterns from hull geometry and speed^[1]. The model was trained on 420 inviscid free-surface simulations generated for two parent yacht hulls, each with 70 variants evaluated at three speeds. On a geometry-held-out test set, ShipNet achieved R^2 scores of 0.98 for hull pressure and 0.91 for wave fields. Inference requires approximately 0.15 seconds per case. Meanwhile, a separate study proposed a novel methodological framework for monitoring floating anthropogenic debris in urban rivers using fixed, in-situ cameras and deep learning techniques^[2]. This framework includes a geometric model to estimate the actual size of detected objects from 2D images.