Prediction of Ultra-High-Performance Concrete (UHPC) Compressive Strength Based on Convolutional Neural Networks

This paper investigates the potential of deep learning in predicting the compressive strength of ultra-high-performance concrete (UHPC) by developing a convolutional neural network (CNN) model with two convolutional layers. The proposed CNN architecture demonstrates capability in accurately predicting UHPC compressive strength from tabular data encompassing various material compositions. Ten input variables were selected, including the cement content, water content, silica fume content, silica powder content, silica sand content, superplasticizer content, and curing parameters. The model was trained and tested on a dataset comprising 219 samples. Experimental results indicated excellent predictive performance, with the CNN achieving a coefficient of determination (R²) of 0.959 on the test set and a mean absolute percentage error (MAPE) of 5.55%, demonstrating both accuracy and stability. Comparative analysis revealed that the CNN's performance was comparable to established machine learning methods like XGBoost (R² = 0.961), which are typically more suited for tabular data. Furthermore, SHAP (SHapley Additive exPlanations) analysis confirmed the model's interpretability. These findings collectively suggest that the CNN-based approach shows considerable promise for predicting compressive strength across diverse UHPC formulations.