Data-Driven Predimensioning: Applying Graph Neural Networks to Reinforced Concrete Design

Predimensioning in structural engineering has traditionally been an experience-driven process that relies heavily on the skills of the engineer and general formulas. While machine learning methods, in particular Graph Neural Networks (GNNs), have been used for optimization in various engineering domains, their application to the predimensioning of reinforced concrete (RC) structures has been notably absent. This study fills this gap by developing a synthetic dataset specifically focused on the predimensioning of RC columns and slabs. The dataset is rigorously simulated using Finite Element Method (FEM) software, with reinforcement calculations that comply with deflection and concrete crack width limits closely aligned with Eurocode standards. The generated data serve as an ideal basis for training, validation, and testing of GNNs. The primary goal is to automate the selection of optimal cross-sections for these RC elements. In doing so, this study pioneers the application of GNNs to the specific challenges of predimensioning in RC structures, thereby introducing a data-driven approach to this critical aspect of structural design.