With this project, we aim to verify a new theoretical approach to speed-up hyper-parameter search for a hydrodynamic force predictor based on a machine learning approach. The dataset for this training has been generated using high-fidelity simulations of floating structures which target applications for offshore renewables (floating offshore wind, wave power), using state-of-the-art software. The implemented hyperparameter search is based on a Bayesian optimizer to be run over tens of data, network, and physical parameters, which will then identify the best network configuration. Also, to reduce the dependency of this search over the randomicity of the initial guesses, per each network, this search is repeated 20 times. Critical to our research objective, is the definition of the loss functions for this algorithm, which comprises two main metrics: the model accuracy on the physical system, and the physical stability of the network output. The latter is part of ongoing research of this project members, and for which an original and novel procedure is employed. Its scope is to curtail the training time and bound it by the total training time rather than on the response reconstruction via physical integration.