Predicting the compressive strength of UHPC with coarse aggregates in the context of machine learning

Abstract

Compared to traditional UHPC, UHPC-CA exhibits higher compressive strength and elastic modulus, but also renders the prediction of compressive strength more complicated. In order to reduce the error and training time, a Machine Learning method is investigated to predict the compressive strength of UHPC-CA, including aggregate strength for the first time. Firstly, a series of cubic compressive strength experiments are conducted to testify the difference between UHPC-CA and traditional UHPC, which also provide 35 data points for the UHPC-CA database. Additionally, 213 data points are added to the database from the literature review to complete the dataset. This database was employed to train and validate six ML models: BR, SVM, MLP-ANN, DT, KNN and RF. The findings suggest that all the employed models demonstrate satisfactory predictive capability regarding the cubic compressive strength of UHPC-CA. Notably, the RF model outperforms others, exhibiting an R2 value of 0.853. Feature sensitivity analysis identified the volume fraction of steel fibers, water reducer amount, and silica fume content as the three most influential factors affecting the compressive strength of UHPC-CA. Finally, the interaction among the content of aggregate, the strength of aggregate, and the volume fraction of steel fibers was analyzed. The recommended inclusion levels for the aggregate range from 0.3 to 0.6, for the fiber from 2% to 2.5%, with a minimum aggregate strength of 150 MPa.

Keywords

• Compressive strength
• Feature sensitivity analysis
• Machine learning
• Multi-parameter pattern prediction
• RF model
• UHPC-CA