Multi-Objective Optimization of Expanded Polystyrene Concrete Using Machine Learning and Nonlinear Programming

This study presents a comprehensive framework integrating machine learning models with genetic algorithm-based optimization for multi-objective design of Expanded Polystyrene Concrete (EPSC), addressing the critical challenge of balancing lightweight benefits with mechanical performance requirements. Four ML algorithms—Random Forest (RF), Artificial Neural Network (ANN), Support Vector Regression (SVR), and Gaussian Process Regression (GPR)—were evaluated using 5-fold cross-validation on 120 experimental data points to predict compressive strength, flexural strength, impermeability, and density. RF and GPR demonstrated superior performance, particularly for density prediction (R²=1.0, RMSE≈0) due to deterministic mass-balance relationships, while compressive strength prediction achieved moderate accuracy (GPR: R²=0.3736, RMSE=3.3978 MPa, ~14% relative error). However, flexural strength and impermeability proved challenging (R²≤0.11), highlighting their sensitivity to microstructural factors beyond compositional inputs. Multi-objective optimization using MATLAB's gamultiobj genetic algorithm with weighted objectives (compressive strength: 0.4, flexural strength: 0.3, impermeability: 0.2, density: 0.1) identified 150 Pareto-optimal solutions after 150,000 function evaluations. The optimal design, corresponding precisely to the EPSC-25 experimental category (0.365 kg/m³ EPS, 25% replacement), achieved simultaneous improvements: 9.9% compressive strength enhancement (28.90 vs. 26.30 MPa), 4.1% density reduction (2175.43 kg/m³), 2.8% impermeability improvement, and unchanged flexural strength. This non-monotonic behavior validates that optimal EPS levels exist where microstructural benefits—improved pore structure distribution, reduced shrinkage cracking, and enhanced particle packing efficiency—outweigh inherent strength penalties. The framework reduces experimental burden by 80-90% compared to traditional trial-and-error approaches, providing engineers with diverse trade-off solutions for applications ranging from high-strength structural elements (28-29 MPa at 2175-2180 kg/m³) to maximum lightweight components. This data-driven methodology advances sustainable EPSC design through efficient computational optimization integrated with limited experimental validation, demonstrating that even ML models with moderate predictive power can guide effective optimization when properly integrated within multi-objective frameworks.