Physicochemical-aware ensemble learning and design of sustainable concrete

Machine learning approaches have been used to accelerate the design of sustainable concrete, but conventional machine learning models lack generalizability and interpretability, making it challenging to design concrete with solid wastes that involve different physicochemical properties. This paper presents a physicochemical-aware ensemble learning approach for designing concrete and demonstrates generalizability for diverse wastes with varying physicochemical properties. The approach integrates concrete knowledge into a prediction-optimization framework that combines ensemble learning models and multi-objective optimization. The framework explicitly considers the physicochemical characteristics of raw ingredients and utilizes a knowledge graph to reveal mechanistic insights into how wastes influence concrete properties. The results revealed that physicochemical-aware ensemble learning enabled the consideration of diverse wastes. Ultra-high-performance concrete mixtures were designed, reducing material cost by 39% and carbon footprint by 61%, while retaining high compressive strength (>120 MPa). The approach has the potential for developing sustainable concrete while maximizing waste valorization.