Pre-trained machine learning for inverse structural design of piecewise developable surface

Abstract

This paper addressed the challenge of inverse design in structural engineering, focusing on predicting reinforcement and thickness parameters for piecewise developable reinforced concrete shells. Specifically, it investigates whether pre-trained machine learning models can more effectively predict rebar directions and thicknesses from displacement data compared to models trained from scratch. To answer this question, large datasets were used to pre-train two ML models for rebar direction and thickness prediction, which were then fine-tuned on a small dataset representing a specific shell geometry. The results show that pre-training ML significantly improves prediction accuracy and efficiency for the thickness task, while offering moderate computational benefits for the rebar direction task. This finding is important for structural engineers and computational designers seeking fast, data-efficient workflows. The work paves the way for future research on integrating geometric information and developing scalable, domain-specific pre-training strategies for structural design problems.