Content area
Full text
Title No. 119-S103
This study uses machine learning (ML) to characterize the shear strength of reinforced concrete (RC) beams and one-way slabs. A database of 1436 RC shear tests is used to train three ML algorithms (ordinary linear regression, support vector regression, and Gaussian process regression). The database is divided into two training subsets per ACI 318-19 minimum shear reinforcement requirements. Each training sample consists of up to 10 predictive features (geometry, materials, reinforcement detail, load location) and the corresponding target feature (shear strength). The most accurate algorithm, the Gaussian process regression, has 17% and 11 % smaller mean percent error in shear strength predictions than ACI 318-19 guidelines for the subsets with shear reinforcement areas less than and larger than the minimum ACI 318-19 requirement, respectively. This algorithm also provides predictive confidence with a standard deviations less than 5.2 kN (1.18 kip) for 99.7% of the predictions.
Keywords: artificial intelligence; influencing parameters; regression; shear capacity; slabs.
INTRODUCTION
This paper explores an application of machine learning (ML) in structural engineering-more specifically, in predicting the shear strength of reinforced concrete (RC) beams and one-way slabs. The focus is on shear because accurately predicting the shear strength of RC members has been more challenging compared to predicting the flexural strength. Data generated over the years through hundreds of experiments provide insight into key shear-resisting mechanisms such as shear stress in the concrete compression zone, aggregate interlock, dowel action of longitudinal reinforcement, and stirrups (Collins et al. 2008). This study was motivated by the fact that these test data also provide an opportunity to employ ML methods for shear strength predictions.
Unlike ML, which can consider all parameters simultaneously, for practical reasons, existing studies on the prediction of shear strength (Angelakos et al. 2001; Bazant et al. 2007; Kim et al. 2012; Tureyen and Frosch 2003) focus only on select parameters they evaluate to be the most influential at the time of the study. However, the set of influential parameters can change when new research and data reveal additional influential parameters. For example, ACI 318-19 (ACI Committee 318 2019) recently included the effect of member size and the amount of longitudinal reinforcement for predicting shear strength, as documented by Moehle (2019). There is no doubt that ACI...