Interpretable Machine Learning for Prediction of Post-Fire Self-Healing of Concrete

Area of Science:

• Civil Engineering
• Materials Science
• Computational Science

Background:

• Accurate prediction of concrete's self-healing is crucial for material engineers and contractors.
• Machine learning (ML) offers high-precision predictions but is often complex ('black box').
• Limited research exists on ML applications for cementitious material self-healing.

Purpose of the Study:

• To pioneer the use of ML for predicting post-fire self-healing in concrete.
• To compare the performance of four ML methods (SVM, RT, ANN, ET) for this prediction task.
• To identify key factors influencing concrete's self-healing capacity.

Main Methods:

• A comprehensive database was compiled from literature.
• Twelve input variables (e.g., w/c ratio, temperature, curing) were analyzed.
• Four ML models were optimized and compared; Monte Carlo analysis validated model stability.

Main Results:

• All tested ML models showed superior precision compared to linear regression.
• The Ensemble of Regression Trees (ET) model exhibited the highest accuracy and robustness.
• Model interpretation identified temperature, curing regime, and aggregate content as key predictors.

Conclusions:

• ML, particularly the ET model, provides accurate and interpretable predictions for concrete's post-fire self-healing.
• This approach enhances understanding and forecasting of concrete repair capabilities.
• Key factors influencing self-healing were identified, aiding material design and engineering.