Accuracy of Neural Networks for Surrogate Modelling of the Sag of Hanging Cables

Abstract

The creation of surrogate models is a classical problem in Machine Learning. The present paper is a case study of training a surrogate model for a real-life engineering problem: the computation of the sag of a cable hanging between two pylons. Neural networks have been trained using samples of the solution for several physical parameters. A parametric study of the role of three hyperparameters (the number of training samples, the size of the network and the initialization of gradient descent) is presented.

Full document