Semi-supervised learning method incorporating structural optimization for shear-wall structure design using small and long-tailed datasets

Abstract

Intelligent structural design based on machine learning represents a novel structural design paradigm and has received extensive attention in recent years. However, the performance of the machine learning models is heavily dependent on the quality and quantity of training data, as the underlying approaches are inherently data-driven. Well-recognized data issues ‒ particularly data insufficiencies and long-tailed data distributions ‒ have become critical impediments in this research area. To address these data issues, this study formulates a schematic structural design task as a semi-supervised learning problem. Specifically, a semi-supervised learning method using small, long-tailed datasets is proposed in which a structural optimization method is incorporated into a self-training framework. As a practical application of the proposed method, a shear-wall layout optimization procedure is devised, based on a two-stage evaluation strategy and the previously established empirical design rules. Results of the numerical experiments indicate that the proposed method can improve the design performance on the tail data by 11.6 % and reduce the performance difference between the head and tail data by 21.3 %, compared to the conventional supervised learning method. A typical case study shows that the shear-wall layout created by the proposed method can satisfactorily resemble that by design engineers whilst meeting key code-specified requirements.