Graph-based deep learning for compressive behavior of 3D-printed concrete columns across layers

Extrusion-based three-dimensional concrete printing enables automated, formwork-free fabrication, but anisotropy and weak interlayer bonding complicate reliable prediction of the mechanical behavior of 3D-printed concrete (3DPC). This study developed a framework employing two parallel approaches, the finite element method (FEM) and physics-informed neural networks (PINNs), to simulate the compressive response of 3DPC. An open-source FEM formulation incorporating node-to-surface contact mechanics and the Mazars damage model was first established and validated against an experimentally tested 3DPC column, yielding a 19.5% underestimation of load capacity and displacement errors within 7.7%. PINNs were then adopted to embed governing mechanical laws directly into learning, improving physical consistency and interpretability over purely data-driven prediction. Two architectures, a multilayer perceptron (MLP) and a novel residual graph neural network (GNN), were evaluated against the FEM results. The GNN achieved a lower displacement root mean square error of 0.002 mm than the MLP (0.003 mm) while requiring fewer trainable parameters (90 vs 141). Benign generalization analysis indicated that graph convolution reduced the required signal-to-noise ratio by approximately 1.4×, thereby enabling the model to extract the true signal in the presence of greater noise. Explainability analyses further showed that the z -coordinate dominated compressive-response prediction, while tangential coordinates governed interlayer behavior, and subgraph visualization revealed physically consistent stress and deformation pathways. The results show that coupling physics constraints with graph learning enhances both accuracy and interpretability, providing a pathway toward reliable structural design in 3DPC construction.