Abstract Reinforced concrete (RC) members subjected to seismic loading experience variable axial forces and shear spans, which induce axial–shear–flexural interaction that governs their hysteretic behavior. Existing interaction models rely on complex constitutive assumptions such as smeared crack formulations, resulting in high computational demand and convergence difficulties, limiting their applicability for system-level assessments. To address these limitations, this study proposes a machine learning (ML)-based parameter adaptation approach that embeds ML surrogate models into lumped plasticity formulations with an adaptive updating mechanism. The suggested framework enables skeleton and hysteretic parameters to be dynamically updated during the analysis in response to variations in axial load, shear span, and cumulative damage, thereby capturing both cyclic degradation and axial–shear–flexure interaction. The approach is implemented for RC columns through the development of an adaptive hysteretic model. Ensemble neural networks, optimized via a complexity-aware grid search, are trained on a database of 286 cyclic tests with a balanced distribution of flexure-, flexure–shear-, and shear-critical specimens. These surrogate models are integrated into OpenSees through cross-platform coupling of Python-based ML routines and C++ numerical analysis to develop the adaptive model. Examination of adaptive features, coupled with experimental validations, confirm the accuracy of the proposed approach in capturing hysteretic axial–shear–flexural interaction under varying axial loads and shear spans, while achieving superior predictive performance and markedly reduced computational demand compared with existing interaction models. The results show that neglecting axial–shear–flexural interaction may significantly overestimate deformation and energy dissipation capacities. To this end, the proposed approach is shown to provide a reliable and efficient method for incorporating these effects into seismic resilience assessments.
Wu et al. (Mon,) studied this question.