Machine Learning for Estimation of Dynamic Model Parameters of Autonomous Underwater Vehicles: A Review for 2015-2025 Period

The operational capability of Autonomous Underwater Vehicles (AUVs) depends on the precise modeling of their dynamic behaviors under environmental disturbances. Traditionally, model parameter estimation processes—conducted through tank tests, empirical calculations, and Computational Fluid Dynamics (CFD) analyses—are giving way to data-driven approaches due to high costs, intensive computational loads, and real-time adaptation constraints. This study systematically reviews the machine learning (ML) techniques developed for estimating AUV dynamic model parameters over the ten-year period from 2015 to 2025. Within the scope of this review, classical methods such as Support Vector Regression (SVR), Artificial Neural Networks (ANN), and Multi-Output Gaussian Processes (MOGP) are examined alongside Physics-Informed Neural Networks (PINN), which integrate physical laws into the learning process, and Explainable Artificial Intelligence (XAI) approaches that ensure model transparency. Furthermore, LSTM and Transformer architectures, which model the temporal dependencies of AUV motions, and Reinforcement Learning (RL) based online adaptation strategies are analyzed. The reviewed methods are presented comparatively in terms of data requirements, computational complexity, physical consistency and validation strategies. This review serves as a guide for researchers in the field of underwater robotics, highlighting the current state and future research directions of modern machine learning paradigms in AUV system identification processes.