Data-Driven Machine Learning Informed Maneuvering and Control Simulation

A data-driven machine learning (ML) informed strategy has been implemented to enhance reduced-order models in maneuvering and control simulation for BB2 - a variant of the Joubert submarine model. The ML-informed approach for maneuvering and control simulation was developed by the Naval Surface Warfare Center - Carderock Division (NSWC-CD) as part of the Hydrodynamic Digital Twins (HDT) project. The ML models for BB2 were trained based on the free running model (FRM) test data collected by the Maritime Research Institute Netherlands (MARIN). The training dataset consists of ship motion data from various maneuvers such as vertical and horizontal overshoots and controlled turns. A genetic algorithm (GA) was employed to optimize the neural network (NN) architecture and its hyperparameters during the training process. These trained NN models were integrated into Maneuvering and Control Simulation (MCSIM) to correct predictions of total hydrodynamic force and moment, thereby enhancing maneuvering predictions for BB2. The performance of both the NN models and ML-enhanced MCSIM was evaluated through a series of validation and generalization runs.