This project took place in the class "Computational Intelligence in Engineering" by Prof. Bernd Markert and assisted by Arnd Koeppe and Marion Mundt at the Institute of General Mechanics at RWTH Aachen University. The Code was written in a team consisting of Vivek Chavan (Github: @Vivek9Chavan), Chi Shing Li (Github: @charles4444) and myself.
Summary: An artificial neural network that predicts boundary forces on a 2D cube, based on displacement and stress. The model was trained by a preexisting FEM dataset.
The goal of this project was to test if neural networks are sufficient enough to be used with or instead of classical FEM simulations. While these have long known issues, such as massive computational cost and a loss of information and memory with the slightest change in the input parameters, neural networks can be trained on already existing data to reduce the time and energy load on future simulations. In this case a boundary value problem is looked on, using stress and displacement values on the edge of a 9 x 9 node cube. The goal is to predict the force distribution within the cube and compare these to the results of the FEM solution. Two different models have been built. For one a convolutional neural network, inspired by AlexNet, aswell as a custom built model with LSTM layers.
Architecture of the CNN:
The project showed some promising results, such as the CNN solution shows:
I specially want to thank Arnd Koeppe and Marion Mundt for assisting in the project and want to redirect interested readers to their research for further reading.
Koeppe, A., Barmer, F., Markert,B., Model reduction and submodelling using neural networks, 2016, Proc. Appl. Math. Mech. (PAMM) 16, 537-538
Koeppe, A., Barmer, F., Markert,B., An intelligent nonlinear meta element for elastoplastic continua: deep learning using a new Time-distributed Residual U-Net architecture, 2020, Computer Methods in Applied Mechanics and Engineering 366, 113088
Koeppe, A., Barmer, F., Markert,B., Mundt, M., Hesser, D.F., Mechanik 4.0. Künstliche Intelligenz zur Analyse mechanischer Systeme, 2020, Handbuch Industrie 4.0: Recht, Technik, Gesellschaft, 553-567