Data-driven machine learning forecasting and design models for the tensile stress-strain response of UHPC

The tensile behavior of ultra-high performance concrete (UHPC) is distinctive from conventional concrete (CC) and is typically included in design. This study leverages machine learning (ML) techniques and regression analysis to predict the full stress-strain behavior of UHPC in tension and characterizes the effects of various parameters. A comprehensive database comprising 500 data points from 24 experimental programs was assembled. Multiple ensemble learning algorithms, including gradient boosting, extreme gradient boosting (XGBoost), bagging regressor, and extremely randomized trees (ExtraTrees), were evaluated and compared against traditional multiple linear regression. The ExtraTrees model outperformed others, resulting in a mean absolute error (MAE) of 0.524 MPa, a root mean square error (RMSE) of 1.140 MPa, and a coefficient of determination (R2) of 0.80 for predicting the first cracking stress (FCS). For strain at first cracking (SFC), ExtraTrees achieved an MAE of 0.0175 mu epsilon and an R2 of 0.953. The model also performed well for post-cracking stress (PCS) and post-cracking strain (SPC). Of the six feature inputs considered -compressive strength (fc') of UHPC matrix, fiber reinforcement index (RI), fiber length (Lf), fiber volume (Vf), fiber diameter (Df), and fiber type-interpretation by SHAP values revealed that fc', Vf, and RI are the most influential. Additionally, a webbased interface was developed using the ML models, allowing users to predict FCS, SFC, PCS, and SPC and generate a bilinear tensile stress-strain curve, applicable for both softening and hardening UHPC types. The interface is intended for design purposes.