Thanks for your questions.
First of all, I do not agree the opinion of explanation totally. Generally speaking, the neural network or machine learning is explainable mathematically, though there are really some gaps to get the physical meaning for specified application. Neural network is just a mathematical tool for simulate a non-linear relationship, the fuzzy characteristic is inherent to some extent. If there is some confusion of understanding explanation of NN, I think you could treat it as a method of system identification. It is maybe helpful. So, if we consider the application in a mathematical way, for example, wave prediction in my paper, it is one of the best ways to solve the time-series prediction.
The reason for me to choose NN for my application can be shown in two aspects.
From the view of neural network, in 2019-2020, so many IT and technology company (Facebook, Apple, Amazon), institution and university (MIT, Carnegie Mellon University) have applied neural network, especially LSTM RNN into their productions for time-series problems. With the selection of these successful application (Speech Reignition, Context Reignition, Intelligent Assistance) is also a solid support for my chosen algorithm. Also, there is paper to illustrate wave spectrum parameter prediction with NN. There is a promising way to solve engineering issues with the development of IT and computational technology.
From the view of wave prediction application, to solve the problem of non-linearity and non-causality as I mentioned in my slides, a functional method needs to be applied. Each neural network has their own advantages. I have just selected the best suitable for my application. For my application, there are several methods to do the wave prediction, like physical-principle method, system identification. They also have good result if there is enough computational time. However, what condition I want to apply the algorithm is online real-time control. So considering the paradox between efficiency and accuracy of feedforward control process, I need a algorithm with have a simple architecture and small computational time. The principle-based method has large simulation time because of a complicated model structure. So, I have chosen NN as the prediction method in my application. Another advantage of NN is that the neural is not fixed by the physical model. The NN model can be replaced by a new training process at any moment with new data, but some physical-principle algorithm is based on the single condition. For online real time control, the NN have significant flexibility if there is change of WEC device in the operational condition. So the NN algorithm is suitable for my wave prediction.
In conclusion, although traditional physical model has good accuracy, the fast NN method is more suitable for this application when considering the real-time control. The algorithm selection is based on the advantage of NN and the need of wave prediction application.
If there is any further discussion, please do not hesitate to send it to me.