Transformer-Based Surrogate Model for Nonlinear Response of Steel and Steel-Fiber Reinforced Concrete Structures
Please login to view abstract download link
The reliable prediction of nonlinear, history-dependent responses in reinforced concrete (RC) and steel-fiber reinforced concrete (SFRC) structures is crucial for structural design and performance assessment. This work presents a Transformer-based scientific machine learning framework trained on high-fidelity nonlinear finite element simulations capturing concrete cracking, fiber- and rebar–concrete interaction, fiber bridging effects, and steel elasto-plasticity. By leveraging the self-attention mechanism, the Transformer learns long-range temporal dependencies across the entire loading history, enabling continuous prediction of structural responses such as load–displacement evolution and maximum crack width. Unlike feedforward neural networks limited to single load states or recurrent models that struggle with long-range dependencies and parallelization, the Transformer efficiently maps full loading paths while maintaining high computational efficiency. The framework is applied to RC and hybrid RC–SFRC beams, demonstrating accurate reproduction of complex temporal phenomena governed by cracking, bond deterioration, and material nonlinearity. A global sensitivity analysis illustrates how the learned temporal representations can be used to investigate parameter influence over the full load path. The Transformer framework can be extended to multi-span beams, prestressed elements, and diverse load combinations, aiming to provide a generalizable, real-time surrogate for concrete structural design and performance prediction under complex loading scenarios. This approach lays the groundwork for future digital decision-support tools and accelerated virtual testing of nonlinear structural systems.
