A data-driven approach to improve model uncertainty of concrete crack prediction in determining SLS target reliability.

Abstract

This paper reports a data-driven approach using Artificial Neural Network (ANN) machine learning tools to predict crack widths in reinforced concrete due to irreversible serviceability limit state (SLS) load-induced cracking. SLS target reliability levels in design standards such as Eurocode and those of South Africa were assigned using ultimate limit state values. Where SLS cracking is the dominant criterion, these levels are insufficient, needing a full probabilistic analysis. With SLS cracking the limiting criterion in the design of reinforced concrete water retaining structures and bridges, these types of structures would benefit from improvements to both crack prediction and suitable reliability levels. The semi-analytical SLS load-induced crack formulations in design standards have a model uncertainty CoV in the order of 0,35 to 0,38, significant in probabilistic analysis and reliability (where general structural uncertainty CoV is 0,1. Model uncertainty as a random variable is highly dependent on the crack formulation considered, making target reliability assessment challenging. The ANN model aims to improve crack model uncertainty. A dataset compiled from experimental research on load-induced cracking is used to train the ANN model.