Formulation of Drying Shrinkage in Concrete Based on Interpretation of Machine Learning Model

Abstract

This study proposes a new approach to interpret and extract dominant regimes from a machine learning model for constructing formulas, and provides a case study of the regression task for drying shrinkage in concrete. The effects of the concrete mix and aggregate properties on drying shrinkage were extracted based on the proposed approach. The results showed complex regimes, such as the non-linear effect of the mixing ratio based on the unit water content and aggregate volume, and the effect of the water absorption of aggregate, which varies with the mixing ratio. These findings were compiled into a formula and it was confirmed that the predicted value calculated by the formula tracks actual shrinkage well and the regression performance is improved especially in regions where data density is low.