Title:
Machine Learning-Based Optimization of Ultra-High-Performance Concrete for Flexural Strength and Sustainability (Prepublished)
Author(s):
Fayez Moutassem
Publication:
Materials Journal
Volume:
Issue:
Appears on pages(s):
Keywords:
Bayesian neural networks; CO2 emission; flexural strength prediction; multi-objective optimization; SHAP, sustainability; ultra-high-performance concrete
DOI:
10.14359/51749415
Date:
12/18/2025
Abstract:
This study presents a machine learning–driven framework for the sustainable design of ultra-high-performance concrete (UHPC) mixtures with a focus on maximizing flexural strength while minimizing material cost and embodied CO₂ emissions. A curated dataset of 333 UHPC mixtures was developed, incorporating 13 input features including binder composition, steel fiber dosage, and curing parameters. A Bayesian Neural Network (BNN) was trained to predict flexural strength with high accuracy (R² = 0.936, RMSE = 1.37 MPa, MAE = 1.09 MPa), supported by residual analysis confirming minimal prediction bias and robust generalization. SHAP analysis was used to interpret model predictions and identify key drivers of flexural behavior—namely, curing time, steel fiber dosage, and silica fume content. The BNN was coupled with the NSGA-III algorithm to perform multi-objective optimization and generate Pareto-optimal UHPC mixtures. A utility-based scoring method was introduced to select designs aligned with different project priorities—enabling the identification of fiber-rich, high-strength mixtures as well as low-emission, cost-efficient alternatives. The framework supports field-level implementation and is well-suited for integration with sustainability rating systems such as LEED or Envision. It provides a transparent, generalizable, and industry-ready tool for intelligent UHPC mixture optimization, advancing data-driven design practices for green infrastructure applications.