Neural network-based material modeling for porous materials like aerogel

Kurzfassung

Aerogels and aerogel compositeshave huge potential for industrial applications especially within the automobile and aerospace industry. However, to validate the suitability of a certain aerogel for a specific application, modelling and simulations of a wide range of aerogels becomes imperative to select from. Classical approaches towards the material modelling of the different types of aerogels is time and resource consuming. The objective of this master thesis is to develop a neural network-based material model to predict the stress-strain response of a foam-like material under compressive or tensile loading and scale it for aerogels using reverse engineering approaches. The data required to make predictions for the stress{strain response is generated using Abaqus software. After analysing the different material models available for foam-like materials, the hyperfoam material model was selected for data generation as the stress-strain response generated using the hyperfoam model corresponds closely with the aerogel data (nature of the stress-strain response and material parameters). The dataset used for developing the neural network contains 6000+ simulations generated with Abaqus scripts. The stress strain responses were generated for possible combinations of the material parameters. It was required to correlate the stress-strain series data with discrete values. Since, the loading conditions were considered same (i.e., no impact loading), there were 2 possibilities to combine the discrete and series data { replicating discrete values for all time steps of the series data or using a hybrid architecture of the neural network. Due to the same loading conditions, the discrete values were replicated for each time step of the series data. Replicating discrete values led to an extremely non-linear data which makes it very difficult for any simple feed forward neural network to predict. Hence, the stress strain response was approximated using NumPy Polyfit function by means of a 3rd degree polynomial. Hence, each experiment then resulted in a set of 4 coefficients. A multivariate feed forward neural network was trained to predict these coefficients with an accuracy of over 90%. After training the neural network and reaching the desired accuracy, the network had to be scaled to predict stress-strain response for aerogels. Thus, part of the dataset was used for scaling the network. Correlating the responses predicted by the neural network with the experimental stress-strain response provided important insights for understanding the physical significance of the parameters. The scaled network was later validated for the other subset of the aerogel data. Although the thesis presents an approach to reduce the time for modelling and simulations of aerogels, it can be extended by working on the finite element implementation of the neeural network-based material model.