An Integrated Approach for Design Optimization in AI-Driven CAE

Abstract

Listen

Translate article

Prior to actual testing, performance evaluation is conducted through Computer-Aided Engineering (CAE) in design process. The analysis results are scrutinized to verify whether the design meets the intended performance. If not satisfied, the conditions are modified, and the analysis is iteratively performed until the desired performance is achieved. However, this iterative process poses challenges in terms of both time and cost. To address this, machine learning and deep learning methodologies have been developed. Nevertheless, some of the existing models focus on identifying a single optimal solution for a given performance. This focus may overlook the potential to discover diverse sets of design parameters that can achieve the same performance criteria. In this study, we propose a method that aims to efficiently identify multiple design parameters that satisfy the desired performance with high accuracy, thereby reducing the need for iterative processes. We modify the model to predict various design parameters by applying Monte Carlo dropout and Bayesian neural network to the tandem network, which outputs only one design parameter fitted to the training data. Furthermore, we propose a 2-stage methodology for exploring local minima by performing bayes optimization based on multiple candidate values derived from the tandem network. The proposed model significantly reduces the number of iterations required for design optimization and predicts multiple possible design parameters. Experimental results using data from rear-wheel steering and braking simulations demonstrate the overwhelming performance and diverse design parameters provided by our proposed model.