Durability prediction of sustainable marine concrete under freeze-thaw cycles using multi-objective machine learning models
Marine and cold-region concrete structures are prone to surface scaling, interior microcracking, and long-term strength degradation due to freeze-thaw cycles (FTCs). Understanding and accurately predicting FTC-induced damage is essential for sustainable and resilient structural design. This study in...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-07-01
|
| Series: | Case Studies in Construction Materials |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2214509525005856 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Marine and cold-region concrete structures are prone to surface scaling, interior microcracking, and long-term strength degradation due to freeze-thaw cycles (FTCs). Understanding and accurately predicting FTC-induced damage is essential for sustainable and resilient structural design. This study introduces machine learning (ML)-based algorithms for forecasting FTC-induced durability of concrete while simultaneously predicting sustainability metrics such as CO₂ emissions and production cost. To enhance accessibility and practical usability, a graphical user interface (GUI) was developed, enabling engineers to input concrete mix parameters and obtain performance predictions without requiring programming knowledge. Four machine learning techniques were utilized: a convolutional neural network (CNN), a genetic algorithm with optimized artificial neural network (GA-ANN), an adaptive neuro-fuzzy inference system (ANFIS), and multi-objective optimization (MOO). The dataset used in this study was compiled from previously published experimental results in the literature, allowing the development of predictive models without the need for new laboratory testing. All models demonstrated strong predictive accuracy (validation R² > 0.90), aligning well with observed experimental trends. The MOO model achieved the highest performance with R² training = 0.974, R² testing = 0.96, MAE = 1.12 MPa, and RMSE = 1.54 MPa, outperforming CNN (R² = 0.942 and 0.93), GA-ANN (R² = 0.931 and 0.902), and ANFIS (R² = 0.905 and 0.939) for training and testing phases, respectively. Evaluation metrics included R², MAE, MSE, RMSE, and MARE, with model ranking as follows: MOO > ANFIS > CNN > GA-ANN. The MOO model also achieved the most accurate predictions for CO₂ emissions and production cost (R² = 0.98 and 0.99, respectively). These findings highlight the reliability of ML approaches in predicting concrete performance under FTC conditions, reinforcing their potential for use in the design of durable marine infrastructure. |
|---|---|
| ISSN: | 2214-5095 |