Generic approach to plausibility checks for structural mechanics with deep learning
Year: 2017
Editor: Anja Maier, Stanko Škec, Harrison Kim, Michael Kokkolaras, Josef Oehmen, Georges Fadel, Filippo Salustri, Mike Van der Loos
Author: Spruegel, Tobias; Schröppel, Tina; Wartzack, Sandro
Series: ICED
Institution: Friedrich-Alexander-University Erlangen-Nuremberg, Germany
Section: Resource Sensitive Design, Design Research Applications and Case Studies
Page(s): 299-308
ISBN: 978-1-904670-89-6
ISSN: 2220-4342
Abstract
The simulation of product behavior is a vital part in virtual product development, but currently there is no tool or method available that can examine the quality of FE simulations and decide automatically on whether a simulation is plausible or non-plausible. In the paper a method is presented that enables automatic plausibility checks on basis of empirical simulation datasets. Nodal simulation data is transformed to numerical arrays, of fixed size, using virtual spherical detector surfaces. Afterwards the arrays are used to train a Deep Convolutional Neural Network (AlexNet). The Neural Network can then be used for plausibility checks of FE simulations (structural mechanics). In a first application a Deep Convolutional Neural Network is trained with simulation data of a demonstrator part, the rail of speed inline skates. After the GPU training of the Neural Network, further simulations are evaluated with the net. These simulations were not part of the training data and are used to calculate the prediction quality of the Neural Network. This approach is to support development engineers during design accompanying FEA in virtual product development.
Keywords: Simulation, Computational design methods, Integrated product development, Plausibility checks, Deep learning neural network