Abstract
This paper presents a neural network model for predicting a ship's magnetic field distribution at arbitrary depths and courses using inverse modeling. Trained on synthetic FEM-generated data, the model addresses limitations of the multi-dipole approach, which is sensitive to positional errors and computationally demanding. The data preparation process and Bayesian optimization of network architecture and hyperparameters are described. Model accuracy is assessed using various metrics and dataset sizes, with comparisons to the multi-dipole model in terms of prediction accuracy, computational cost, and robustness. When ship positioning data were disturbed, the relative deterioration of selected quality indices was six to seven times lower than for the multi-dipole model. Although requiring more data for high accuracy, the neural model is faster, more robust, and well-suited for degaussing and risk-assessment applications.