Hybrid data-driven approaches to predicting the compressive strength of ultra-high-performance concrete using SHAP and PDP analyses

Ultra-high-performance concrete (UHPC) is a cutting-edge and advanced constructions material known for its exceptional mechanical properties and durability. Recently, machine learning (ML) methods play a pivotal role in predicting the compressive strength (CS) of UHPC and evaluating the dominant input parameters for a suitable mix design. This research, three hybrid machine learning models were utilized: Random Forest (RF), AdaBoost (AB), and Gradient Boosting (GB) algorithms with particle swarm optimization (PSO), namely AB-PSO, RF-PSO, and GB-PSO, to predict compressive strength and perform SHAP (Shapley additive explanation) analysis. To build predictive hybrid ML models, a dataset of 810 experimental data points was collected for compressive strength (CS) from published literature. Additionally, SHAP interaction plots were generated to visualize the impact of each feature on a specific prediction made by the model. Our results indicate that hybrid machine learning models perform better than traditional models, and the hybrid GB-PSO model showed the high prediction accuracy among models. The hybrid GB-PSO model had higher precision compared to the other two models. It achieved R2 values of 0.9913 during the training stage and 0.9804 during the testing stage for the prediction of CS. The SHAP analysis revealed that age, fiber, cement, silica fume, and superplasticizer had a significant influence on compressive strength, while the impact of other input parameters was comparatively lower. The PDP (Partial Dependence Plots) analysis results amount of individually input variables material can be calculated simply for the designed CS. These findings are valuable for construction applications and offer essential insights for design engineers and builders, aiding their understanding of the significance of each component in UHPC.