What question did this study set out to answer?

The aim is to develop a machine learning framework to predict water penetration depth in seawater-based self-compacting concrete.

March 16, 2026Open Access

Hybrid and explainable machine learning for predicting water penetration depth in seawater based self-compacting concrete

Key Points

The aim is to develop a machine learning framework to predict water penetration depth in seawater-based self-compacting concrete.
Created a dataset of 144 SCC specimens using seawater and varying silica fume and fly ash content.
Evaluated seven machine learning algorithms including decision tree and hybrid gray wolf optimizer stacking model.
Used Z score normalization and tenfold cross-validation for model training and evaluation.
Conducted explainability analyses to identify influential factors affecting water penetration depth.
The hybrid model achieved the highest performance with R² of 0.9568 and RMSE of approximately 1.19 mm.
Silica fume and chloride concentration were identified as key factors influencing water penetration depth.
The model provides meaningful insights for sustainable SCC mixture design while reducing reliance on standard testing.

Abstract

The depth of water penetration under pressure (DWP) is a critical durability indicator for concrete, particularly in marine environments where precise evaluation is essential for predicting long term performance. Due to the high cost, long duration, and limited flexibility of standard DWP tests, this study introduces an advanced machine learning (ML) framework for rapid and reliable prediction of DWP. A unique and rarely reported experimental dataset of 144 self-compacting concrete (SCC) specimens was produced, featuring a combined design that uses seawater both in mixing and curing, together with controlled variations of silica fume (SF) ranging from 4% to 15% by mass of binders and fly ash (FA) ranging from 20% to 31% by mass of binders. This integrated laboratory dataset, which has been scarcely explored in previous research, enables a comprehensive assessment of permeability behavior under aggressive conditions. Seven ML algorithms including decision tree (DT), random forest (RF), extra tree regressor (ETR), extreme gradient boosting (XGboost), gaussian process regression (GPR), artificial neural network (ANN), and a newly developed hybrid gray wolf optimizer (GWO) Stacking model were evaluated using Z score normalization and tenfold cross validation. Although advanced individual models exhibited strong predictive capability, the hybrid ensemble achieved the highest performance with coefficient of determination (R²) equal to 0.9568, root mean square error (RMSE) of approximately 1.19 mm, and an Index of agreement (IA) value of 0.9887, indicating excellent accuracy and generalization stability. Explainability analyses based on shapley additive explanations, permutation feature importance (PFI), and accumulated local effects (ALE) revealed that SF and chloride concentration are the most influential factors affecting DWP, whereas FA and sulfate variables show milder and more linear effects. Beyond numerical accuracy, the proposed hybrid model provides meaningful engineering insight by clarifying the influence of key mix design parameters including cement content, aggregate proportions, superplasticizer dosage, and the water to powder ratio. This interpretability supports the development of sustainable and performance based SCC mixtures and offers an efficient and practical alternative to labor intensive laboratory testing for assessing concrete permeability in marine environments.

Hybrid and explainable machine learning for predicting water penetration depth in seawater based self-compacting concrete

Key Points

Abstract

Cite This Study