Research on an intelligent prediction method for the stress field in perforated concrete slabs based on convolutional neural networks
Year: 2026
Type: report
Abstract: Finite-element (FE) analysis is the standard for stress evaluation in perforated concrete slabs, yet its run-time grows rapidly with geometric complexity. This study develops a convolutional neural-network (CNN) surrogate that predicts slab stress fields directly from geometry. Training targets are generated from a parametric FE model that enumerates hole layouts, including symmetric and asymmetric patterns, under shrinkage and service load combinations. To broaden geometric coverage, the dataset is expanded via generative augmentation (DCGAN) and procedural perturbations of hole geometry and spacing. On held-out geometries, the CNN reproduces FE stress with a Pearson correlation of 0,908 and an RMSE of 0,058 MPa for symmetric layouts, while delivering more than 50× faster inference than FE. Case studies on multi-hole slabs indicate that the surrogate remains accurate near re-entrant corners and narrow ligaments where stress gradients are steep, enabling rapid “what-if” exploration and efficient crack-resistance sizing. The proposed FE-supervised learning framework provides a direct and consistent comparison between the FE baseline and the neural surrogate, offering a practical path to real-time design support for large slabs with complex perforations.
Authors: Nan Mou, Zheltkovich, Andrei
Institution: Brest State Technical University
Language: English
Subfield: Building and Construction
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7214697189