Neural networks are able to approximate nonlinear functions with accuracy. This makes them suited for many fields far from their origin in computer science.
In this work we will apply these neural networks to composite materials from two different perspectives: top-down, for structural health monitoring, and bottom-up, for efficient multiscale simulations. The term structural health monitoring refers to the knowledge, usually in real time, of the integrity of a component. In this work we will approach to this problem in a computational manner, by the detection of defects in a composite panel. For this purpose, by means of a grid of virtual sensors in the simulation, we will generate synthetic data from which we will train a neural network that will be able to predict those defects in size and position.
The second problem we will deal with is the use of surrogate models for multiscale simulations. Customarily, the way in which we introduce information from smaller scales into the scale of interest is by means of constitutive laws; the drawback of this is that the accuracy tends to be limited by the difficulty of modelling certain phenomena and characteristics.
The objective of surrogate models is to replace these constitutive laws with a data-driven model that retains an accuracy comparable to direct simulation but at a fraction of the computational cost. To demonstrate the concept, we developed a neural network which is able to reproduce the behaviour of an RVE consisting of carbon fibres embedded in an elastoplastic matrix.