Inverse machine learning for the design of perforated beams: Parent section and material prediction

Abstract

This study introduces a novel machine learning-based inverse design methodology for predicting the parent cross-sectional and material properties of perforated steel beams with elliptically-based openings. Unlike conventional forward design or optimisation-based methods, the proposed approach frames structural design as an inverse problem. It enables the direct mapping of opening geometry and resistance parameters to essential properties such as section depth, flange width, web thickness, flange thickness, and yield strength. Five advanced supervised machine learning models including Histogram-Based Gradient Boosting, Random Forest, Extreme Gradient Boosting, k-Nearest Neighbors, and Support Vector Regression were trained on a dataset generated through forward structural analysis. This methodology develops non-iterative surrogate models that enhance the generalizability and accessibility of the structural design process. The models demonstrated excellent predictive performance, with coefficient of determination (R²) values exceeding 0.99 for most outputs. Shapley Additive Explanations (SHAP) analysis identified web-post buckling resistance and section height as the most influential input features, with other variables contributing depending on the output. The proposed inverse learning framework was also benchmarked against an analytical design model to assess accuracy and consistency. To support practical implementation, a user-friendly web tool was developed, allowing engineers and researchers to access instant predictions. Overall, this research offers an efficient and explainable data-driven solution for structural design, demonstrating the practical value of artificial intelligence in engineering applications.