Generalizable ML Solutions for Predicting Equivalent Viscous Damping Ratio of Diagonal RC Coupling Beams

Diagonally reinforced concrete coupling beams (DRCBs) play an important role in coupled shear-wall systems for high-rise buildings, acting as primary energy dissipators under seismic loading. In current practice, their seismic design depends on adequate shear capacity and detailing—especially diagonal reinforcement and end confinement—to ensure sufficient ductility (rotation capacity) and energy dissipation. However, in recent years, the equivalent viscous damping ratio (EVDR) of coupling beams has not been systematically studied, and its use in simplified assessment remains under-documented. With this motivation, this study compiles a database of 125 DRCBs from experimental programs conducted between 1980 and 2025. We first performed a sensitivity analysis by deriving feature weights over 100 random trials and then evaluated performance using feature subsets ordered from highest to lowest weight. The analysis indicates that \( \left({f}_c^{\prime },{f}_{yd},{f}_{yl},{p}_d,{p}_v,{p}_l\right) \) and aspect ratio (AR) are the most influential predictors of EVDR. Two machine learning (ML) solutions were developed—a single-learner multilayer perceptron (MLP) and an ensemble extreme gradient boosting (XGBoost) model. XGBoost achieved superior accuracy (R2 = 0.98, RMSE = 0.008, MAE = 0.005) and exhibited minimal bias (ME = 0.001 and σME = 0.008; overall balanced grade (OBG) = 0.007). Results show that XGBoost offers the best robustness and generalization for EVDR prediction. A practical solution (PS) was also derived using a minimal MLP structure to provide an explicit equation for practitioners. For validation, a blind test using two specimens from the literature was conducted; predictions from both trained models closely matched the experimental EVDR, demonstrating the generalization and reliability of the developed solutions for real-world applications.