A physics-guided neural network framework for special segment section prediction and automated preliminary design of special truss moment frames

Abstract Special truss moment frames are ductile steel structural systems with excellent energy dissipation capacity. Their design follows the capacity design principle, in which members outside the special segment remain essentially elastic while plastic deformation is concentrated within the special segment through chord-member yielding. This paper proposes a physics-guided neural network framework for special segment section prediction in special truss moment frames and develops a front-end web-based prediction platform for preliminary member selection based on special-segment design parameters. The proposed framework predicts the section properties of the special segment, including the chord-member plastic section modulus, chord-member moment of inertia, vertical-member plastic section modulus, and vertical-member moment of inertia. Engineering design equations and an expected-shear-strength consistency constraint are incorporated into the learning process to integrate data-driven prediction with mechanical consistency. To evaluate in-distribution predictive performance while avoiding augmentation-induced data leakage, the original 154 floor-level samples were divided before data augmentation. Gaussian-noise augmentation was applied only to the training samples, increasing the 123 original training samples to 492 samples, while the 31 test samples were retained in their original, non-augmented form. The expected-shear-strength consistency assessment further showed that the predicted section-property combinations maintained reasonable consistency with the mechanical relationships required by capacity design. The feasibility of simplified automated preliminary member selection was further demonstrated through a 4-story, 45-ft-span special truss moment frame example. Additionally, a lightweight web-based application was developed using Streamlit for real-time prediction of section properties and expected maximum shear strength through interactive parameter input, facilitating rapid comparison of design alternatives at the preliminary design stage. The proposed framework provides a physically guided design-support approach for special-segment section-property prediction and simplified automated preliminary member selection in special truss moment frame design.