A physically compliant surrogate constitutive model for thin-walled rods using DNN - design, training and use in FEM simulations

Resumo

We propose a deep neural network (DNN)-based surrogate constitutive model tailored for thin-walled rod members. Unlike traditional multilayer perceptrons, our architecture is physically informed by design, embedding key mechanical behaviors directly into its structure. By formulating the internal energy function in terms of conventional rod strain measures, the model provides a compact yet robust representation of complex structural responses in a low-order (rod) context. The data generation workflow, training strategy and modelling choices are outlined, highlighting the essential fine-tuning required for accurate performance. Once trained, the surrogate model is integrated into our in-house finite element software, PEFSYS, which specializes in large displacement and finite strain analyses of rods and shells simulation. The model accurately captures key nonlinear effects, including stiffness degradation due to local flange and web buckling, offering an efficient and physically consistent framework for simulating thin-walled structures.