A NEURAL NETWORK APPROACH FOR COMPRESSIVE STRENGTH PREDICTION OF ULTRA-HIGH-PERFORMANCE CONCRETE: DEVELOPMENT, ANALYSIS, AND ROBUSTNESS TESTING

Abstract

The development of Ultra-High-Performance Concrete (UHPC) is constrained by the empirical and costly nature of mixture optimization, necessitating robust predictive models that capture the complex, non-linear interactions between its constituent materials. While Artificial Neural Networks (ANNs) have been widely adopted for this purpose, their application is frequently criticized for lack of transparency, inconsistent interpretability, and uncertain generalization beyond training datasets. This study addresses these critical gaps by developing, analysing, and rigorously validating a fully connected ANN model to predict UHPC compressive strength using a substantial and curated dataset of 810 mixture designs. The model, featuring a single hidden layer with 100 neurons and L2 regularization, demonstrated high predictive accuracy on an independent test set, achieving an R² of 0.922, RMSE of 11.267 MPa, and a mean absolute percentage error (MAPE) of 8.305%. Twenty-fold cross-validation confirmed robustness, revealing consistent performance across most folds while identifying specific regions in the mixture design space where predictive accuracy diminished, highlighting data limitations. A critical, multi-method feature importance analysis—employing Univariate Regression, RRelieff, and SHapley Additive exPlanations (SHAP)—provided a nuanced hierarchy of influential variables. While filter methods ranked silica fume and fiber content highest, SHAP analysis revealed the ANN’s internal logic prioritizes curing age, cement content, and fiber dosage as the most dominant predictors, aligning model reasoning with fundamental materials science principles. This work moves beyond a mere performance report to deliver a critical evaluation of ANN utility in UHPC science, explicitly confronting challenges of model interpretability, context-dependent feature significance, and generalization. It concludes that while ANNs are powerful predictors, their effective translation into reliable design tools necessitates larger, more diverse datasets, standardized benchmarking against alternative algorithms, and the integration of explainable AI techniques to yield actionable, physically meaningful insights for engineers.