Hybrid generative-ensemble approach for predicting recycled aggregate concrete strength properties.

Area of Science:

• Civil Engineering
• Materials Science
• Sustainable Construction

Background:

• Recycled aggregate concrete (RAC) is crucial for sustainable construction.
• Predicting RAC's mechanical properties from mix proportions is challenging due to data limitations.
• Accurate property prediction is vital for reliable structural applications.

Purpose of the Study:

• To develop a hybrid generative-ensemble framework for predicting key mechanical properties of RAC.
• To address data scarcity issues in RAC mix design.
• To identify dominant mix proportion variables influencing RAC properties.

Main Methods:

• A conditional variational autoencoder (CVAE) was employed to generate synthetic data, augmenting a database of 112 RAC mixes.
• Seven supervised learning algorithms, including gradient boosting and support vector regression, were trained and validated.
• Feature-attribution analysis was conducted to determine the influence of mix proportions on mechanical properties.

Main Results:

• The hybrid framework accurately predicted compressive strength, split tensile strength, flexural strength, and elastic modulus.
• Gradient boosting and support vector regression demonstrated superior predictive performance and stability compared to baseline models.
• Binder-related variables were identified as primary drivers of strength, while aggregate-related variables influenced stiffness.

Conclusions:

• The proposed framework offers a data-driven approach for efficient screening of RAC mixes.
• The findings provide interpretable insights for optimizing sustainable RAC mix design.
• This methodology enhances the practical application of RAC in construction.