Joint Eurocode-Compliance Classification and Reinforcement Regression with a Multi-Task Graph Neural Network Surrogate for Reinforced Concrete Predimensioning

Early-stage structural design requires rapid exploration of large design spaces, where the initial sizing of reinforced concrete members shapes downstream material use, cost, and the number of design iterations. Conventional predimensioning relies on experience and simplified formulae, while finite element analysis remains too slow for iterative use. This study presents a multi-task graph neural network surrogate that predicts per-element Eurocode compliance together with the required reinforcement for reinforced concrete slab-and-column buildings in one pass. A shared GraphSAGE encoder, trained on 2562 synthetic building graphs from automated finite element simulations, feeds one head for a compliance probability and another for reinforcement quantities. Because the rule-based Eurocode check is a hard pass-or-fail decision that does not vary smoothly with the design, the surrogate learns a continuous, differentiable compliance probability in its place, demonstrated for two representative criteria, one per element type, namely the l/250 deflection limit for slabs and the 4% reinforcement-ratio limit for columns. Across five random seeds, cost-sensitive focal-loss training that weights missed non-compliance above false alarms reached 90.9% balanced accuracy and held the share of non-compliant elements wrongly passed as compliant at 6.1% for columns and 1.6% for slabs, with a mean reinforcement error near 2% of the normalised target range. Inference averaged approximately 0.5 ms per building, between five and six orders of magnitude faster than the finite element analyses. A differentiable, multi-task graph surrogate therefore supports fast, cost-sensitive compliance screening for early-stage predimensioning, serving as a seed for gradient-based design exploration and a starting point for finite element verification.