Estimating RC corrosion distribution from surface cracks using mesoscale analysis integrated with machine learning

Abstract

Understanding the degree of reinforcing bar corrosion in reinforced concrete (RC) structures is crucial for evaluating their residual performance. This study proposes a simulation system for estimating the distribution of corrosion along the rebar of a RC beam member based on surface crack widths. The system integrates the rigid body spring model (RBSM) with machine learning methods. The inputs are surface crack widths and the desired output is the distribution of corrosion-induced expansion. A large dataset of training samples for machine learning is generated by running RBSM simulations using different expansion distributions. After training with this dataset, the neural network is able to correlate inputs and outputs, allowing it to estimate an expansion distribution from given cracking data. The estimated expansion distribution is then used to simulate the surface cracks using RBSM, and the error between the given (input) cracking data and simulated cracks is returned as an input to the trained network in order to optimize the results and enhance performance of the system. The applicability of this RBSM-neural network system is validated using both synthetic and experimental test data. The estimation results correlate well with the target data, demonstrating the effectiveness of the system in estimating internal expansive strain along the rebar and reproducing the cracking distribution using surface crack data. Internal distributions of cracking and stress are extracted from the simulations, providing additional information for analyzing structural performance.