Thin-Walled Compliant Mechanism Component Design Assisted by Machine Learning and Multiple Surrogates

Abstract

This work introduces a new design algorithm to optimize progressively folding thin-walled structures and in order to improve automotive crashworthiness. The proposed design algorithm is composed of three stages: conceptual thickness distribution, design parameterization, and multi-objective design optimization. The conceptual thickness distribution stage generates an innovative design using a novel one-iteration compliant mechanism approach that triggers progressive folding even on irregular structures under oblique impact. The design parameterization stage optimally segments the conceptual design into a reduced number of clusters using a machine learning K-means algorithm. Finally, the multiobjective design optimization stage finds non-dominated designs of maximum specific energy absorption and minimum peak crushing force. The proposed optimization problem is addressed by a multiobjective genetic algorithm on sequentially updated surrogate models, which are optimally selected from a set of 24 surrogates. The effectiveness of the design algorithm is demonstrated on an S-rail thin-walled structure. The best compromised Pareto design increases specific energy absorption and decreases peak crushing force in the order of 8% and 12%, respectively. Recently, our research group has proposed the use of HCA topometry (thickness distribution) methods for compliant, plastically deformable structures to support the generation of conceptual design for progressively folding thin-walled components including S-rails [28]. This method follows the design principles of compliant mechanisms and their ability to transfer force and displacement from a given input

Optimal Design of Nonlinear Multimaterial Structures for Crashworthiness Using Cluster Analysis

“…While more research is needed, the parametric studies in this work show that the analysis (crash simulation) of the cluster designs is a suitable indicator to approximate the optimal number of clusters. Other approaches such as the elbow method, information criterion approach, and the Silhouette method [51] are computationally less expensive but not as effective as the parametric study shown in this work. Ongoing investigation aims to control the structural complexity of the clustered design generated by the proposed approach.…”

“…As an alternative, metamodels can be derived by sampling the dynamic, nonlinear finite element model. The resulting metamodels are numerically inexpensive and allow to find nearoptimal solutions through the use of global multi-objective algorithms [51]. The key aspect to using metamodels for global optimization lies in balancing between global exploration and local exploitation.…”

“…Several metamodels have been evaluated for design problems in crashworthiness including: polynomial response surface, radial basis functions, and Kriging [51]. Based on cross-validation errors, Kriging is the preferred metamodel and it is used in this work.…”

Cluster-Based Optimization of Cellular Materials and Structures for Crashworthiness

“…The input observations to the cluster analysis algorithm is the conceptual material distribution x * ∈ R n and the desired number of clusters K, where 1 K n; usually, K n (Algorithm 1 [32]). A discussion on the optimal value of K can be found in [33].…”

“…In other words, it is desirable to generate an accurate metamodel that explores a large portion of the design space with a few sampling points [34]. Several metamodels have been evaluated for design problems in crashworthiness, including: polynomial response surface, radial basis functions, and Kriging [33]. Based on cross-validation errors, Kriging is the preferred metamodel and it is used in this work.…”