Predicting critical crack propagation length in sustainable additive-enhanced concrete using explainable machine...

Area of Science:

• Materials Science and Engineering
• Civil Engineering
• Computational Mechanics

Background:

• Predicting critical crack propagation length (CCPL) in sustainable additive-enhanced concrete (SAEC) is crucial for structural durability.
• Existing experimental and numerical methods for CCPL prediction are often complex, costly, and computationally inefficient.
• There is a need for accurate, efficient, and data-driven methods to predict CCPL in SAEC.

Purpose of the Study:

• To develop and evaluate a comprehensive machine learning (ML) framework for predicting the CCPL of SAEC.
• To compare the performance of ensemble, kernel-based, and deep learning models, with a focus on a novel Neural Tangent Kernel Gaussian Process (NTK-GP) model.
• To identify key features influencing CCPL and provide uncertainty quantification for the predictions.

Main Methods:

• Developed an ML framework integrating ensemble, kernel-based (NTK-GP), and deep learning models.
• Utilized a dataset of 800 SAEC samples with nine key features, subjected to controlled experimental testing.
• Evaluated model performance using statistical indices (R², RMSE, MAPE, VAF) and cross-validation; employed Explainable AI (SHAP) for feature importance and interaction analysis; used bootstrap for uncertainty quantification.

Main Results:

• The NTK-GP model demonstrated superior predictive performance (R²=0.95–0.96, RMSE=0.74–0.90 mm, MAPE=0.09–0.14, VAF=0.95–0.96).
• Explainable AI analysis identified fiber type (FT) and fiber volume content (FVC) as the most influential features ( >65% variance), with significant interactions.
• Statistical tests confirmed NTK-GP's comparability to state-of-the-art models, and bootstrap intervals validated predictive reliability.

Conclusions:

• The proposed ML framework, particularly the NTK-GP model, provides an accurate, efficient, and interpretable alternative for CCPL prediction in SAEC.
• The study highlights the significant impact of fiber type and volume on crack propagation, offering insights for material design.
• This data-driven approach advances structural durability analysis and supports the development of sustainable, fracture-resistant concrete materials.