Data-Driven Assessment of Concrete Mixture Compositions on Chloride Transport via Standalone Machine Learning Algorithms

This paper employs a data-driven approach to determine the impact of concrete mixture compositions on the temporal evolution of chloride in concrete structures. This is critical for assessing the service life of civil infrastructure subjected to aggressive environments. The adopted methodology relies on several simple and complex standalone machine learning (ML) algorithms, with the primary objective of establishing confidence in the unbiased prediction of the underlying hidden correlations. The simple algorithms include linear regression (LR), k-nearest neighbors (KNN) regression, and kernel ridge regression (KRR). The complex algorithms entail support vector regression (SVR), Gaussian process regression (GPR), and two families of artificial neural networks, including a feedforward network (multilayer perceptron, MLP) and a gated recurrent unit (GRU). The MLP architecture cannot explicitly handle sequential data, a limitation addressed by the GRU. A comprehensive dataset is considered. The performance of ML algorithms is evaluated, with KRR, GPR, and MLP exhibiting high accuracy. Given the diversity of the adopted concrete mixture proportions, the GRU was unable to accurately reproduce the response in the test set. Further analyses elucidate the contributions of mixture compositions to the temporal evolution of chloride. The results obtained from the GPR model unravel latent correlations through clear and explainable trends. The MLP, SVR, and KRR also provide acceptable estimates of the overall trends. The majority of mixture components exhibit an inverse relation with chloride content, while a few components demonstrate a direct correlation. These findings highlight the potential of surrogate approaches for describing the physical processes involved in chloride ingress and the associated correlations, toward the ultimate goal of enhancing the service life of civil infrastructure.

💡 Research Summary

This paper investigates the use of a variety of standalone machine‑learning (ML) techniques to predict the temporal evolution of chloride concentration in concrete, based on concrete mixture proportions and environmental conditions. The authors compiled a comprehensive dataset from published experimental studies, encompassing 12 mixture‑related input features (water, SPRC, OPC, water‑to‑binder ratio, fly ash, silica fume, GGBS, super‑plasticizer, fine and coarse aggregates) together with exposure variables (surface chloride level, exposure time, temperature, depth). After normalizing the data (zero mean, unit variance) the set was split into 75 % training and 25 % testing subsets, and a 10‑fold cross‑validation scheme was applied to mitigate over‑fitting.

Seven ML models were evaluated: three relatively simple regressors—linear regression (LR), k‑nearest‑neighbors (KNN), and kernel ridge regression (KRR)—and four more sophisticated approaches—support vector regression (SVR), Gaussian process regression (GPR), a multilayer perceptron (MLP), and a gated recurrent unit network (GRU). Model performance was quantified using coefficient of determination (R²), mean absolute error (MAE), mean squared error (MSE), root‑MSE, and mean absolute percentage error (MAPE).

Results show that LR performed poorly (R² ≈ 0.62), reflecting the strongly nonlinear nature of chloride ingress. KNN achieved moderate success (training R² ≈ 0.89, test R² ≈ 0.82). Both KRR and SVR delivered high accuracy (test R² ≈ 0.90 and 0.89, respectively). GPR attained the best overall score (test R² ≈ 0.91) while also providing predictive uncertainty estimates through its Bayesian framework. The MLP, despite lacking explicit temporal handling, achieved a comparable test R² of 0.90. In contrast, the GRU network struggled (test R² ≈ 0.55), likely due to the limited size and heterogeneity of the dataset, which impeded learning of sequential dependencies. Consequently, the GRU was excluded from subsequent sensitivity analyses.

A detailed sensitivity study was conducted by fixing a reference concrete mix (water 184 kg/m³, OPC 460 kg/m³, fly ash 100 kg/m³, fine aggregate 700 kg/m³, coarse aggregate 1050 kg/m³, super‑plasticizer 1.8 kg/m³, exposed to 19.6 g/L surface chloride for 1.3 years at 9 °C) and varying each input variable across its observed range. The influence of water content illustrated divergent model behaviours: LR, KNN, and MLP suggested a direct relationship (higher water → higher chloride over time), whereas KRR, SVR, and GPR indicated an inverse relationship, which aligns better with physical expectations that excess water increases porosity and thus facilitates ion transport. Similar analyses for other constituents revealed that most mixture components (e.g., supplementary cementitious materials, super‑plasticizer) tend to reduce chloride ingress, while a few (notably water) can have a positive effect. Notably, increasing the amount of sulfate‑resisting Portland cement (SPRC) consistently lowered predicted chloride levels across all high‑performing models (KRR, SVR, GPR, MLP), supporting the known durability benefits of SPRC.

The study demonstrates that data‑driven surrogate models can capture complex, multivariate relationships governing chloride transport with far lower computational cost than physics‑based diffusion or finite‑element simulations. However, the poor performance of the GRU underscores the necessity for larger, well‑structured time‑series datasets when employing recurrent architectures. Future work is suggested to expand the dataset, explore hyperparameter optimization, integrate ensemble techniques, and incorporate uncertainty quantification to further enhance reliability for service‑life prediction of concrete infrastructure under aggressive environments.