Machine learning for strength prediction and optimal design of sustainable concrete formulas

Abstract

Given the large environmental impact of the concrete industry, which represents 8- 9% of global CO₂ emissions, the design of concrete mixes with low carbon footprints that still meet structural performance requirements will be an essential part of global decarbonization efforts. In this work, we build a concrete performance model, which maps from concrete constituents to compressive strength, a key structural property. Specifically, we leverage the quantity and quality of information provided by our industrial concrete partners (whereas most existing related studies use small, narrow datasets derived from laboratory experiments) to establish an improved concrete performance model that captures the role of several concrete ingredients and a wide variety of formulas. We find that the features which are predicted to be important to concrete strength are compatible with industry knowledge, and that predictions can be improved in the case of small datasets by leveraging information from other larger datasets. Additionally, we integrate our machine learning model into an optimization procedure, and identify mixtures which have minimal cost and minimal climate impact. Lastly, we discuss the trade-offs between these two design parameters, and how these considerations differ by the required strength of the concrete.