Researchers have developed optimized machine learning models to estimate the compressive strength of concrete exposed to sanitary wastewater. The study found that the best-performing model, PSO-XGBoost, achieved a coefficient of determination (R²) of 0.96 on test data, demonstrating the potential of artificial intelligence to support concrete strength assessment in wastewater infrastructure.
Machine Learning Approaches for Concrete Strength Prediction
Predicting wastewater-exposed concrete strength presents challenges because of the complex, nonlinear relationships between material properties and environmental conditions. To address this, researchers combined particle swarm optimization (PSO) with four machine learning algorithms: XGBoost, gene expression programming (GEP), support vector regression (SVR), and LightGBM. Multiple linear regression (MLR) served as a baseline for comparison.
PSO is an optimization technique inspired by collective behavior in nature. It searches for suitable model parameters by moving a population of candidate solutions through a defined search space. In this study, the method was used to identify hyperparameter configurations that minimized validation errors and improved the predictive performance of the selected algorithms.
Dataset Preparation and Model Evaluation
The research team assembled 138 samples from previously published studies on machine learning-based prediction. Input features were normalized to a range of −1 to 1, while outliers representing approximately 5% of the samples were removed using a two-tailed statistical method based on the ±3σ rule.
The prepared dataset was divided into training and testing groups, with 110 samples allocated for training and 28 reserved for independent testing. PSO was then applied to optimize the hyperparameters of XGBoost, GEP, SVR, and LightGBM, including parameters such as learning rate, tree depth, and the number of estimators.
Performance was assessed using the coefficient of determination (R²), mean absolute error (MAE), root-mean-square error (RMSE), and the a-20 index. The latter evaluates how closely predictions fall within a ±20% range of actual values, providing an additional measure of model reliability. The researchers also used Shapley additive explanations (SHAP) to examine how individual input variables influenced predictions.
PSO-XGBoost Achieves the Strongest Results
Among the evaluated approaches, PSO-XGBoost delivered the best overall performance during K-fold cross-validation, recording a mean R² of 0.95 and an RMSE of 1.5 MPa. Its optimized configuration included a learning rate of 0.10, 100 estimators, and a maximum tree depth of five.
The other optimized models also demonstrated predictive capability. PSO-SVR achieved a mean R² of 0.92, while PSO-LightGBM and PSO-GEP each recorded a mean R² of 0.90. The study reported that PSO-XGBoost achieved an R² of 0.97 on training data and 0.96 on testing data, with a test RMSE of 1.25 MPa.
SHAP analysis identified the water-to-cement (W/C) ratio as a major factor influencing wastewater-exposed concrete strength. This interpretability method helps researchers understand the contribution and direction of individual variables rather than relying solely on overall accuracy scores.
The researchers also developed a prototype graphical user interface to demonstrate how the models, particularly PSO-XGBoost, could be used for practical strength estimation. The interface remains a conceptual demonstration rather than a fully established field application.
Comparisons with findings from other studies indicated improvements of approximately 10–15% in R² and 15–25% in RMSE for the optimized models. However, differences in datasets and evaluation procedures limit direct comparisons. The relatively small dataset also remains a constraint on the models’ generalizability.
Future Scope for Wastewater Infrastructure Assessment
The findings indicate that optimized machine learning could help estimate reductions in concrete compressive strength under sanitary wastewater exposure. Nevertheless, broader validation is needed before the models can be relied upon across different structures, materials, and operating environments.
Future research could expand the dataset to include concrete mixture composition, exposure duration, hydrogen sulphide (H₂S) concentration, pH, and temperature. Incorporating these variables could improve generalizability and support more direct predictions of residual strength and structural deterioration.