Geopolymer Concrete Crack Prediction System

The transition to sustainable construction, utilizing low-carbon materials such as geopolymer concrete (GPC) reinforced with non-corrosive fiber-reinforced polymer (FRP) bars, necessitates advanced, quantitative structural health monitoring (SHM). Current automated crack inspection often relies on traditional machine learning (ML) classification models (e.g., SVM), which, while achieving high accuracy in categorizing failure modes, inherently fail to provide the quantitative parameters (crack width, length, and area) essential for engineering assessment and maintenance prioritization. To address this critical utility gap, this study proposes an enhanced Deep Learning Semantic Segmentation framework: an Attention-based U-Net architecture. This model is specifically designed with a residual encoder and optimized with a hybrid Dice and Focal loss function to counteract extreme class imbalance and enhance the detection of fine, hairline microcracks characteristic of fiber-bridged GPC systems. The framework achieves high segmentation fidelity, evidenced by a mean Intersection over Union (mIoU) score ranging from 85%–95% on complex GPC/FRP crack patterns. This pixel-level accuracy enables the robust post-processing extraction of maximum crack width (Wmax) and total crack length (Ltotal). This methodological shift from qualitative classification to verifiable quantitative segmentation provides the necessary empirical foundation to track damage evolution, assess serviceability limits, and inform predictive maintenance schedules for novel GPC/FRP composites where standard structural codes are currently lacking.