Beyond Traditional Computer-Aided Design Parameterization, Feature Engineering for Improved Surrogate Modeling in Engineering Design

Artikel i vetenskaplig tidskrift, 2025

Surrogate modeling in engineering design uses Computer-Aided Design (CAD) to create input features. To this end, CAD models are parameterized and have traditionally assisted design changes, automation, and standardization. However, this process leads to low flexibility, limited design space exploration, a high-dimensional design space, and ultimately extended design cycles. This paper builds on an existing methodology of correlation-based
feature extraction in CAD to prevent dimensionality excess and improve the flexibility of surrogate models. We extend the ’sleeping parameters’ concept from extraction to engineered features and position it in the overall machine modeling learning process. To count for efficacy validation as part of the process of training a prediction model, several correlation matrices are suggested to rank and select these new features, which complete the feature
engineering loop. Utilizing a new case study on Thin-Walled Beams (TWBs) crashworthiness, we showcase how to construct the medial axis of a beam cross-section and extract numerous features in several categories. The results show meaningful relationships between the sleeping parameters and their resulting crashworthiness outputs. The implications of the findings suggest the possibility of achieving better predictions with fewer parameters and reduced dependency on CAD parameterization, potentially leading to accelerated design iterations in the development of TWBs