Machine-learning based seismic damage limit state models for corroded low-ductility RC buildings

Abstract

Corrosion-induced deterioration significantly compromises the seismic performance of reinforced concrete (RC) frame buildings, and the risk is compounded for low-ductility buildings constructed prior to the introduction of seismic design codes, yet robust tools for predicting their performance levels remain limited. This study proposes a five-step machine learning (ML)-based framework to predict damage state thresholds for corroded low-ductility RC frames. A finite element (FE) model capable of considering variability in material, geometric, and non-uniform corrosion deterioration is developed. An extensive dataset is generated using Quasi-random sequence, covering a broad range of structural, material, and deterioration parameters. Nonlinear static pushover analyses are conducted to monitor component-level damages and map into global-level responses in terms of maximum inter-story drift (MIDR) limits corresponding to five damage states. The damage states correspond to yielding of reinforcement, spalling of cover concrete, crack formation in joints, and shear failure of RC members. An Artificial Neural Network (ANN) regression model, optimized using Bayesian hyperparameter tuning, is trained and evaluated using five-fold cross-validation. The model achieved R² values ranging from 0.81 to 0.98, demonstrating strong and consistent predictive accuracy across all damage states. To facilitate practical adoption, a Python-based graphical user interface (GUI) is developed, enabling engineers to obtain real-time DS threshold predictions without requiring expertise in ML or FE modeling. The proposed framework offers a computationally efficient alternative to conventional methods for seismic performance assessment of aging RC structures.