Kelvin wakes are fluid motions generated by a moving disturbance at a free surface. We present a machine learning-based framework for inferring the properties of such moving disturbances from the Kelvin-wake patterns. We perform phase-resolved simulations to establish a dataset of nearly half a million Kelvin wakes generated by disturbances of varying propagating speed, length scale and geometry. Trained with the augmented data, the neural network achieves accuracies of 99.7% and 92.4% in predicting the velocity and the length scale of the disturbance, respectively, even if a random noise has been added to the training data. The explainability of the neural network is demonstrated by quantifying the contribution of the input data to the prediction, which shows a strong connection with the diverging and transverse waves. The accuracy of the neural network in predicting the disturbance length scale is sensitive to wave nonlinearity.
