Designing a high-speed craft for better seakeeping in waves can contribute significantly to higher safety and human comfort. Early in the design process, mathematical models such as the 2D+T method are commonly used, while high-fidelity computational fluid dynamics (CFD) and experimental models are used later in the process. Some of the limitations of such models are that they are not fast enough to be used in the ship's system for real-time monitoring or to develop a digital twin. Recently, machine learning methods have demonstrated great promise in building surrogate models from data. These methods include deep learning and recurrent neural network (RNN). In this paper, a systematic investigation of the network architectures and the used optimizers to train the network is presented. Adam, Adagrad, RMSprob and SGD are investigated in training the network. To train the model almost 35000 data points were collected for Fridsma hull operating in 18 regular waves using a 2D+T model. The result showed that gated recurrent unit (GRU) outperformed long short-term memory (LSTM) and RNN in predicting the heave motion. Also, one hidden layer with 5 neurons was enough to achieve mean absolute error of 0.000298 and to predict unseen waves when trained with more than 24000 data points.