Learning functionals via LSTM neural networks for predicting vessel dynamics in extreme sea states

Area of Science:

• Naval hydrodynamics
• Computational fluid dynamics
• Machine learning

Background:

• Predicting vessel motions in extreme sea states is computationally intensive due to complex nonlinear wave-body interactions.
• Existing methods require substantial computational resources, limiting real-time applications.

Purpose of the Study:

• To develop a computationally efficient simulation paradigm for predicting vessel dynamics in severe sea conditions.
• To explore the application of recurrent neural networks (RNNs) for real-time prediction of ship motions.

Main Methods:

• Trained Long Short-Term Memory (LSTM) neural networks (NNs) using data from expensive Computational Fluid Dynamics (CFD) simulations.
• Compared the performance of standard RNNs, Gated Recurrent Units (GRUs), and LSTMs.
• Input: stochastic wave elevation; Output: vessel motions (pitch, heave, roll).

Main Results:

• LSTM neural networks demonstrated superior performance compared to standard RNNs and GRUs.
• Accurate predictions of vessel motions were achieved for a catamaran (sea state 1) and a battleship (sea state 8) in unseen wave conditions.
• Online prediction of vessel dynamics was obtained in a fraction of a second after offline training.

Conclusions:

• The proposed LSTM-based simulation paradigm offers a highly accurate and computationally efficient solution for predicting vessel motions in extreme sea states.
• This work represents the first implementation of the universal approximation theorem for functionals in realistic naval engineering problems.