Using neural networks for dimensioning and certification of mechanical systems: model precision and accuracy

Abstract

Neural networks show impressive performance in lots of domains to handle problems of high complexity. They are universal approximators and can, in principle, be used to learn any type of model. Their use would be of a great benefit as they can be intended to automate major tasks within an engineering project (such as system dimensioning, certification, and criteria verification). However, it is not yet customary to use these technics for lack of competitiveness against forward engineering calculations. One major issue is the robustness and the difficulty to ensure high precisions for deterministic predictions. In this work, we investigate the ability of neural networks to be used to approximate engineering models and their performance in terms of precision and accuracy per target relative error. Increasing accuracy requires understanding how these models work in a deeper way. Applications on use-cases of mechanical structures are used to understand the behaviour of neural networks for this type of problems and illustrate the encountered constraints.