Data-driven prediction and intelligent optimization of strength, porosity and cost of concrete with supplementary cementitious materials

Abstract

The pursuit of sustainable construction has accelerated the use of supplementary cementitious materials (SCMs) in concrete to reduce its environmental impact while ensuring structural performance. However, designing optimal SCM-based mixes involves balancing strength, durability, and cost, which is posing a nonlinear optimization challenge. This study integrates machine learning with multi-objective optimization to predict the compressive strength and porosity and provide practically feasible mix proportions maximizing strength and minimizing porosity and material cost. A dataset of 300 SCM-based concrete mixes was compiled from literature. Random Forest (RF), multilayer perceptron (MLP), Support Vector Regression (SVR), and Ridge Regression models were developed using differential evolution-based hyperparameter tuning. RF achieved the highest accuracy for strength prediction (R2 = 0.97), and MLP for porosity (R2 = 0.96). The developed model also outperformed empirical models in prediction of porosity. The SHAP analysis identified curing time and cement content as the most influential features. The best models were integrated into a Multi-Objective Differential Evolution Optimizer (MODEO) to generate mix designs that minimize porosity and cost while maintaining structural strength. A graphical user interface (GUI) was developed for practical applications. This framework supports performance-based concrete design and offers a robust tool for optimizing structural materials in engineering practice.

Topics & Concepts