A review on Deep Learning Applications for the Simulation of Medium- and Large-Scale Structures

Deep Learning (DL) models have been increasingly adopted in Computational Solid Mechanics to reduce the computational cost of simulation tasks traditionally performed by the Finite Element Method (FEM) such as field prediction, nonlinear or dynamic analysis, and topology optimization. However, it remains unclear how they overcome high-dimensional input encoding and data demand while preserving physical consistency at real-world structural scale (thousands to millions of degrees of freedom). To address this gap, 50 peer-reviewed studies on medium- and large-scale structural simulation are analyzed. The review classifies first-generation DL architectures—Multilayer Perceptron, Convolutional Neural Networks, Generative Adversarial Networks, Recurrent Neural Networks, and Transformers—by input representation, physics integration, and scalability, together with emerging mechanics-oriented frameworks (Graph Neural Networks, Fourier Neural Operators, and Deep Operator Networks), and performs a cross-evaluation based on generalization capacity, data efficiency, and suitability for nonlinear regimes. Evidence shows that purely data-driven models exhibit limited cross-geometry generalization and strong dependence on fixed discretization and large datasets. Physics-informed and FEM-embedded hybrids improve stability and data efficiency, yet scalability often relies on reduced-order targets rather than full-field prediction. Mechanics-oriented frameworks introduce stronger inductive biases that mitigate grid dependence, although validation remains restricted to limited geometric variability. Finally, three priorities for large-scale deployment emerge: stronger cross-geometry transfer, improved reliability and interpretability, and systematic data-generation protocols.