Abstract
To predict the effect of chemical-freezing coupling erosion on the properties of four kinds of FRP-reinforced concrete, rapid freeze-thaw tests were conducted. The mass loss rate, relative dynamic elastic modulus, compressive strength, and flexural capacity were tested to investigate the mechanical performance of specimens. The compression specimens are cylindrical specimens wrapped with FRP, and the flexural specimens are pasted with FRP prismatic specimens on the pre-cracked side. A database was built based on 45 groups of experimental test results, and the prediction effect of the BP neural network and CNN model on compressive strength and flexural capacity was compared. The results showed that CNN performed significantly better. Finally, the maximum number of freeze-thaw cycles was predicted based on the CNN model, providing a new perspective for predicting the durability of FRP-reinforced concrete.
Introduction
Concrete infrastructure is highly susceptible to environmental degradation, particularly in cold regions where freezing temperatures and chemical erosion (from de-icing salts or industrial runoff) occur simultaneously. Fiber Reinforced Polymer (FRP) sheets—including carbon (CFRP), basalt (BFRP), glass (GFRP), and aramid (AFRP)—are frequently employed to reinforce or repair these structures. However, the complex, non-linear interaction of chemical-freezing coupling erosion makes predicting the long-term mechanical performance of FRP-reinforced concrete challenging. This study investigates these degradation patterns and evaluates the efficacy of machine learning models in forecasting structural integrity.
Methodology
The experimental phase involved 168 specimens subjected to up to 100 rapid freeze-thaw cycles while submerged in three distinct chemical environments: 5% sulfuric acid, 2 mol/L sodium hydroxide (alkali), and 3.5% sodium chloride (salt).
Researchers measured mass loss, Relative Dynamic Elastic Modulus (RDEM), compressive strength, and flexural capacity. The specimens included cylindrical forms for compression testing and prismatic forms for flexural testing.
A database of 45 experimental groups was established to train and test two neural network architectures: a standard Back-Propagation (BP) network and a Convolutional Neural Network (CNN). Inputs included FRP properties, environmental stressors, and intermediate degradation metrics (mass loss, RDEM).
Results and Discussion
Environmental Impact
Acid-freeze coupling was identified as the most destructive environment, leading to the highest mass loss and most significant reduction in structural capacity. Interestingly, some specimens exhibited a temporary mass gain during the first 50 cycles due to water absorption and the formation of expansive products like gypsum within the concrete pores, before subsequent spalling led to sharp mass declines. Among the materials, CFRP demonstrated the highest resistance to acid corrosion, whereas GFRP was the most vulnerable.
Model Performance
The CNN model significantly outperformed the traditional BP network in predicting mechanical properties. For compressive strength, the CNN achieved an $R^2$ of 0.945 (vs. 0.823 for BP). For flexural capacity, the CNN reached an $R^2$ of 0.941 (vs. 0.791 for BP). The average relative errors for the CNN were 9.4% for compression and 4.5% for flexure, demonstrating its robustness even with a relatively small dataset.
Conclusion
This study demonstrates that CNNs are highly effective at capturing the complex feature relationships inherent in material degradation under coupled environmental stressors. While the current model is based on 45 experimental groups, future work involving larger datasets and microscopic analysis (SEM/XRD) will further refine these predictive capabilities. The proposed CNN-based method offers a reliable tool for engineers to forecast the durability and lifespan of FRP-reinforced concrete in extreme climates.
Summary updated