Concrete is a complex mixture of dozens to hundreds of ingredients. Conventional concrete is typically a mix of cement, construction aggregate, water, and performance-enhancing chemicals known as admixtures. There are many possible materials that can be used as aggregate, including gravel, crushed stone, sand, slag, recycled concrete, and geosynthetic aggregates of different types and sizes. Since cement is the most carbon-intensive ingredient in concrete, one potential way to reduce the carbon footprint of concrete is to reduce the amount of cement and/or replace it with other cementitious materials like fly ash, slag, and ground glass, which have a lower carbon intensity.
But getting the proportions of cement substitutes right is tricky when we are also trying to adhere to stringent technical performance requirements. Manually optimizing a concrete formula for both sustainable and technically sound outcomes for different end uses is a formidable challenge, one that can be time-intensive. Mathematically, each of the potentially hundreds of possible ingredients is an added dimension in a complex space of all possible concrete formulas. To efficiently and quickly optimize concrete for sustainable performance, and to generate new concrete formulas, we’ve turned to AI.
Accelerating concrete formula discovery with AI
AI can be used to learn and optimize for specific outcomes within high dimensional spaces, where data sets with many attributes can be modeled. When a valid data set is available, AI can be used to estimate or “learn” the feasible high-dimensional space in terms of relevant parameters, such as strength and sustainability. To develop this type of model, we leveraged the work of researchers at the University of Illinois at Urbana-Champaign, including AI expert Prof. Lav Varshney from the electrical and computer engineering department, and concrete expert Prof. Nishant Garg from the civil engineering department.
Varshney and Garg trained a model using the Concrete Compressive Strength data set, openly available from the UCI Machine Learning Repository. This database has 1,030 instances of concrete formulas and their validated attributes, including seven-day and 28-day compressive strength data (i.e., how the concrete gained strength seven days and 28 days after pouring). The embodied carbon footprint associated with the concrete formulas was derived using the Cement Sustainability Initiative’s Environmental Product Declaration (EPD) tool. EPDs are a standardized way of accounting for the environmental impacts of a product or material, including carbon emissions over its life cycle.
Using the input data on concrete formulas along with their corresponding compressive strength and carbon footprint, the AI model was able to generate a number of promising new concrete mixes that could meet our stated data center requirements with a lower embodied carbon impact than the industry standard.