Integrated deep retrofitting framework for progressive collapse control in RC frames using physics infused graph and multi-agent learning process

Abstract

Progressive collapse is an issue in reinforced concrete (RC) frames and it is one of the most substantial threats concerning safety and economy under extreme loading conditions or partial damage scenarios. Heuristic-based methods also fail to generalise over different frame geometries and loading histories in the process. Addressing these shortcomings is done by a completely deep hybrid framework, which consists of five new modules. First, a Spatio-Temporal Graph Attention Network (ST-GAT) models the damages as they occur through the superstructure topology while using dynamic attention weights to find nodes that are most vulnerable with 92.8% efficiency. Second, Conditional Variational Autoencoder for Retrofit Design (CVAE-RD) performs inverse design of retrofitting layouts conditioned on collapse resistance targets, achieving 97.5% valid design rate and 28.4% cost savings. Third, a Physics Informed Neural Network (PINN-RC) acts as mesh-free surrogate model embedding structural equilibrium laws for efficient prediction of collapse responses (R² = 0.987, 85.6% faster than FEM) in the process. Fourth, Multi-Objective Deep Reinforcement Learning agent (MODRL-Retro) learns sequential retrofit policies using Pareto dominating reward vectors, and process improves cost efficiency by 33.9% in process. Finally, a Bayesian Optimization layer (UABO) improves the calibration of retrofitting plans under uncertainty using Gaussian Processes, increasing the confidence in resistance predictions to 95% (± 3%) in process. This integrated framework significantly advances retrofit optimization by capturing the physics, temporal evolution, and uncertainty of collapse phenomena, offering scalable and intelligent retrofitting strategies for real-world RC structures.