Prediction of hysteresis response of steel braces using long Short-Term memory artificial neural networks

• Proposed ANN-based surrogate models for the nonlinear hysteresis response prediction of steel buckling-restrained and conventional hollow structural section braces. • Utilized long short-term memory (LSTM) algorithm for signal-to-signal prediction in proposed surrogate models. • Developed a decoupling technique to overcome the limited experimental datasets. • Validated the steel brace surrogate models using experimental and synthetic numerical data. This article proposes artificial neural networks that utilize the long short-term memory (LSTM) algorithm to estimate the nonlinear hysteresis response of steel buckling-restrained and conventional hollow structural section braces. The proposed models overcome the two main challenges: 1) the complexity of hysteresis response (tensile yielding and strain-hardening in tension, and compressive buckling and strength degradation in compression) and 2) limited training data, using an LSTM network and auxiliary parameters. The development of a suitable training dataset is first presented. The architectures of the proposed models are then described followed by the validation of the model against unseen brace hysteresis responses. The validation results confirm that the proposed LSTM networks are both accurate and computationally efficient in predicting the response of steel braces to random lateral loads, namely axial force – axial deformation response. The proposed models have the potential to be used for seismic response evaluation of steel braced frames, provided that their limitations are properly considered.