An Integrated Prediction Framework for Engineered Cementitious Composite: EDFrame.

Area of Science:

• Civil Engineering
• Materials Science
• Computational Mechanics

Background:

• Engineered Cementitious Composite (ECC) offers superior durability for infrastructure but faces challenges in mixture design due to complex parameter interactions.
• Accurate prediction of ECC mechanical properties, like tensile stress and strain, is crucial for optimizing its application in civil engineering.

Purpose of the Study:

• To develop an integrated prediction framework, EDFrame, for engineered cementitious composite (ECC) performance.
• To enhance ECC mixture design and performance prediction accuracy using advanced machine learning techniques.

Main Methods:

• Collected and curated two original ECC datasets, incorporating 18 features and 10 fiber types.
• Applied a constraints-modified Conditional Tabular Generative Adversarial Network (Tuned-CTGAN) for data augmentation.
• Developed and validated a 1D-Residual Convolutional Neural Network (1D-Residual CNN) for predicting tensile stress and strain, compared against five ML models.

Main Results:

• Tuned-CTGAN significantly improved the 1D-Residual CNN's R-squared values for tensile stress (0.8658 to 0.9128) and strain (0.8433 to 0.9378).
• The EDFrame, incorporating Tuned-CTGAN and 1D-Residual CNN, outperformed all compared machine learning models.
• Partial Dependence Plots (PDP) and Kernel SHAP analyses provided insights into optimal fiber parameters and feature importance.

Conclusions:

• The proposed EDFrame offers a robust and interpretable solution for ECC performance prediction.
• The framework facilitates efficient and accurate mixture design for engineered cementitious composites in practical engineering applications.
• Data augmentation with Tuned-CTGAN is effective in improving the predictive accuracy of ECC mechanical behavior.