Reliability-based mix design for concrete compressive strength using a physics-prior residual-learning surrogate with...

Area of Science:

• Materials Science
• Civil Engineering
• Data Science

Background:

• Predicting concrete compressive strength and designing reliable concrete mixes are crucial in civil engineering.
• Traditional methods struggle under data-scarce conditions, limiting accuracy and efficiency.
• Existing models often lack interpretability and fail to incorporate physical principles effectively.

Purpose of the Study:

• To develop an integrated framework for concrete compressive strength prediction and reliability-based mix design.
• To address challenges posed by data-scarce conditions by combining physics-based knowledge with data-driven approaches.
• To ensure physical plausibility and interpretability in the prediction model.

Main Methods:

• A physics-prior residual surrogate model combining a physics-consistent baseline with a residual neural network.
• Incorporation of effective water-to-binder ratio, age effects, supplementary cementitious material reactivity, and superplasticizer influence.
• Utilized SHAP values and partial dependence curves for model interpretability and feature importance analysis.

Main Results:

• The hybrid Physics+Data model achieved high accuracy (R²=0.9252, RMSE=4.39 MPa) on an independent test set.
• The model maintained accuracy even with only 40% of training samples, demonstrating robustness in data-scarce scenarios.
• Refined uncertainty quantification yielded nominal 95% prediction intervals with ~95.1% empirical coverage.

Conclusions:

• The integrated framework effectively predicts concrete compressive strength and facilitates reliability-based mix design.
• The physics-informed approach ensures physical plausibility and improves model performance, especially under data limitations.
• The study provides a robust tool for optimizing concrete mix design for desired performance targets.