Data-driven prediction of nonlinear sloshing dynamics using POD-integrated and direct LSTM/transformer architectures

Abstract

Accurate prediction of sloshing-induced wall pressures in cryogenic cargo tanks is essential for structural integrity assessment of liquefied natural gas (LNG) and liquid hydrogen (LH2) carriers. This study evaluates sequence-based deep learning architectures for forecasting sloshing pressure fields, incorporating Proper Orthogonal Decomposition (POD) for dimensionality reduction. Pressure-field datasets generated from validated computational fluid dynamics (CFD) simulations are used to train and compare four models (Direct LSTM, POD-LSTM, Direct Transformer, and POD-Transformer) under representative low-energy (smooth) and highenergy (violent, nonlinear) sloshing regimes. The study compares prediction performance and examines the influence of POD at different energy-retention levels. In the low-energy regime, POD improves LSTM accuracy by filtering small-scale fluctuations, whereas the Direct Transformer achieves the highest prediction accuracy using the full pressure field. In the high-energy regime, LSTM-based models fail to reproduce localized impulsive pressures, while the POD-Transformer provides the best balance between prediction accuracy and computational efficiency by suppressing non-dominant fluctuations while preserving dominant impact dynamics. The results demonstrate that POD effectiveness depends on both the sloshing regime and the learning architecture, revealing distinct performance characteristics across reduced-order and full-field sequence-learning approaches.