Virtual engineering technology has undergone rapid progress in recent years and has been widely accepted for commercial product development. Product design and manufacturing organizations are moving from the traditional multiple and serial test cycle approach to simulation, which solves problems and validates performances using CAE and CAD tools.
For an efficient process, it is essential that design variants can be done within a short time frame. This generally leads to a challenge when the system under study exhibits nonlinear behavior. This session introduces a new methodology based on neural networks (NNs) and genetic algorithms (GAs), which “put data to work” and provide the best possible solution for a given design based on the available data. The goal of this methodology is to provide designers with a tool that can be used to select the optimum design for a given product. This is possible thanks to the optimization of the NN itself through GA implementation based on the available training data. Genetic algorithms have been used for neural networks in two main ways: to optimize the network architecture and to train the weights of a fixed architecture.
The performance of a NN is critically dependent on, among other variables, the choice of the processing elements (neurons), the architecture, and the learning algorithm. In particular, the connection density (among neurons) determines its ability to store information and learn from it. On one hand, a reduced number of connections may disable the network to approximate the function. On the other hand, dense connections may cause overfitting. NNs are usually seen as a method to implement complex nonlinear functions using simple elementary units connected with adaptive weights. We focus on optimizing the structure of connectivity for these networks using GAs to reduce learning time and avoid CAD/CAE loops. Indeed, this implementation provides neural network topologies that, in general, perform better than random or fully connected topologies when they learn and classify new data.
Genetic operators, such as mutation and cross-over, introduce variety into the initial randomly connected population, modifying the network’s architecture and testing candidate solutions. Once the most effective NN is trained, it is possible to adjust the design parameters, with the same accuracy as FEA or testing data, but sharply reducing the simulation time: The approximate hour and an half needed to analyze critical points by FEA is reduced to few seconds using neural networks. A MATLAB graphical user interface (GUI) works as a quick design guide, where the training data for the NN is obtained from a set of automatically generated FEA analyses. To assess the effectiveness of this methodology, several practical applications are shown. As an example, the optimal preload for bolted joints is returned in a few seconds starting from bolt’s geometry, friction coefficient, and applied torque.
