In this work, we developed a framework to train DNN-based surrogates for the constitutive equation of thin-walled rods described by kinematically-exact (KE) rod theories. For rod formulations, basic kinematic assumptions are considered, typically limiting cross-sectional motion to a rigid body translation and rotation, along with one extra DOF to account for cross-sectional warping. Despite drastically reducing the total amount of DOFs necessary to model frame structures, local effects such as web/flange bending and buckling are lost. To overcome this, we use energy data from reduced domains of 3D-solid FEM simulations to train a hyperelastic energy function [1], using as arguments the generalized strain measures that arise from the KE rod formulation [2]. Surrogates have been used to capture local effects and higher-order strain interactions in various scenarios [3], [4], [5]. One key aspect of the proposed framework is that a two-step iterative training scheme is adopted alongside special DNN architectures so that the surrogate complies with mechanical constraints: objectivity, positiveness, smoothness, convexity w.r.t. input arguments and energy/stress-free undeformed state. The framework has been implemented and tested for a given I-shape cross-section, with the rod assumed of Simo-Ciarlet’s material with elastic properties resembling steel in a planar frame case. Benchmark studies displayed excellent adherence to reference solutions in cases where traditional KE formulations overestimate stiffness and critical loads, due to their inability to capture softening from local effects. Our solution can reach highly deformed configurations, even when 3D-solid FEM solvers fail to converge. REFERENCES [1] M. P. Kassab, E. M. B. Campello, A. Ibrahimbegovic, A DNN-based surrogate constitutive equation for geometrically exact thin walled rod members, Computation, Vol.13, 2025. [2] M. P. Kassab, E. M. Campello, P. M. Pimenta, Advances on kinematically exact rod models for thin-walled open-section members: Consistent warping function and nonlinear constitutive equation, Computer Methods in Applied Mechanics and Engineering, Vol. 407, 2023. [3] F. As’ad, C. Farhat, A mechanics-informed deep learning framework for data-driven nonlinear viscoelasticity, Computer Methods in Applied Mechanics and Engineering, Vol. 417, 2023. [4] M. Čanađija, V. Košmerl, M. Zlatić, D. Vrtovšnik, N. Munjas, A computational framework for nanotrusses: Input convex neural networks approac