10/21/2025
Researchers from The Grainger College of Engineering have partnered with Meta and Amrize to develop an AI model for optimizing concrete recipes for data centers. Their method aims to improve strength, speed, performance, and sustainability over a manual mix-design approach.
AI models can be a lot of things: a digital personal assistant, an online shopper, and even a travel planner. But what about a chef?
How about a chef that develops recipes not for better-tasting desserts, but for stronger and more durable concrete?
Through a first-of-its-kind collaboration with Meta and cement and concrete producer Amrize, researchers from The Grainger College of Engineering at the University of Illinois Urbana-Champaign are using AI to optimize building materials. By creating custom concrete recipes to align with the specialized needs of AI data centers, their pioneering method improves performance, cuts costs, and reduces carbon footprint compared to traditional concrete mix design.
Concrete is more than a popular construction medium: it’s the most-used material on the planet, next to water. Traditional concrete mixes are made from varying ratios of cement, water, sand, gravel and admixtures. For over 30 years, concrete producers have relied on the same rotation of recipes for all their projects, which vary from residential construction to road paving. However, the recent surge of data centers has introduced new, specialized requirements that require upgrades from traditional mixes.
“These large-scale data centers are hosting much of the data that society is producing, which creates a sense of urgency in building timelines,” said Nishant Garg, an associate professor of civil & environmental engineering and the lead Illinois collaborator of the project. “Their concrete floor slabs must have the high performance and strength-bearing capacity needed to support very expensive GPUs. These data centers are housing the most expensive electronics in the world.”
“The possibilities
are limitless.”
Professor Nishant Garg, Department of Civil and Environmental Engineering
Determining the best concrete recipe for these unique structures is a complicated and time-consuming task. To accelerate the process and improve various metrics like durability and sustainability, Garg’s lab joined forces with Meta to develop a machine learning model that can predict the most optimized concrete recipe based on each developer’s location and priorities. Members of Garg’s lab trained the model on more than 100 unique recipes of mortar and concrete mixes.
“Everything was done in-house,” Garg said. “We received our materials from Amrize and mixed the concrete using a set of starter recipes. We would feed that data into our model, which would then suggest new recipes. We would make those new recipes, gather data on their strength and performance, and feed them back into the model. It's like an iterative loop between humans and AI. Once the model was trained on about 60 concrete mixes, it started to show good performance for the materials from Minnesota. Achieving good predictive performance on this kind of focused data set is one of the key strengths of this AI model.”
Ultimately, the researchers’ model generated a final set of optimized recipes based on the 100-plus original mixes. Following a final check by Amrize, they were pleasantly surprised to find that the AI-optimized mix was 43% stronger in terms of early strength and 35% less carbon intensive than the industry baseline mix. More importantly, it had similar workability for the floor slab application and was cost-competitive to deploy at the large scale. Figure 1 shows a summary of lab results obtained under this collaboration.
The model’s ability to reduce carbon emissions has been a particular source of pride for its creators, as concrete accounts for up to ~8% of all global CO2 emissions. Although concrete itself is an efficient material, the vast amount produced annually results in a large carbon footprint.
One method for reducing carbon in cement production is by replacing CO2-intensive components like Portland cement with waste products or recycled materials, also known as supplementary cementitious materials (SCMs). Examples of SCMs include fly ash — a residual ash produced by coal power plants — and ground granulated blast-furnace slag, a waste product from the iron and steel industry. For producers interested in improving the durability of concrete and lowering its carbon footprint, the novel AI model can suggest concrete recipes sourced from various locally available SCMs.
An optimized recipe generated by the group’s model was recently successfully deployed in a critical section of Meta’s AI data center in Rosemount, MN.
“Data centers can be millions of square feet in size,” Garg said. “It’s a huge amount of concrete being used, and any saving we can do in the cost per cubic yard of the concrete will have massive returns per building. Optimizing even a small percent can save several dollars per cubic yard, which is a sizable saving over the entire project.”
Garg’s lab has built on this success by developing novel methods for speeding up concrete testing, slashing wait times for standard strength tests from weeks to minutes.
“One of the painful things with concrete is how long many of the tests take,” he said. “We have to wait one month for our initial data before designing the next recipe. Our new UR2 test can predict the 28-day performance of SCMs within five minutes. Moving forward, this test will allow us to rapidly iterate among new materials and quickly determine which ones are best for large-scale deployment.”
Garg hopes the cluster of recent advancements will have long-term positive effects on the construction industry. In fact, Meta and the Illinois Grainger engineers have made their model open source to facilitate easier integration.
“We have been focusing on creating open-source tools so that anybody can download and use the model without paying,” Garg said. “We hope that this approach can lead to a large market penetration, and we can help concrete producers that don’t have the budget to license proprietary mix-design software,” he said. “The possibilities are limitless.”
The following articles are available online:
‘Sustainable Concrete via Bayesian Optimization.’ DOI: arXiv:2310.18288
‘UR2: ultra-rapid reactivity test for real-time, low-cost quality control of calcined clays.’ DOI: 10.1016/j.cemconres.2025.107806
Illinois Grainger Engineering Affiliations
Nishant Garg is an Illinois Grainger Engineering associate professor of civil and environmental engineering in the Department of Civil and Environmental Engineering.