At Meta’s Rosemount data-centre project in Minnesota, an AI-guided concrete recipe reached a real construction site. Amrize, Meta, the University of Illinois Urbana-Champaign and contractor Mortenson moved from material tests to a trial slab, then used the mix in a site-support section of a building slab.[1]
Meta reported in March 2026 that the selected mix reached full structural strength 43% faster than the original formula. This concerns strength development, rather than a claim that the finished concrete is 43% stronger.[2]
Concrete has to perform at several ages
A concrete recipe combines a binder, water and aggregate. Replacing some Portland cement with materials such as fly ash or slag can reduce the embodied carbon associated with the binder. But these changes also affect how strength develops. A mix that performs well after four weeks can gain strength more slowly during its first few days.[5]
The earlier research behind BOxCrete modelled that whole strength-development curve. It treated performance at different ages and the carbon footprint as connected design objectives. This allows a search to consider both the early strength needed by a construction schedule and the later strength expected from the finished material.[5]
A model that learns which mix to test next
BOxCrete uses a Gaussian process: a statistical model that predicts an outcome together with uncertainty. Its inputs describe ingredient quantities and curing time. The model can estimate a proposed mixture’s strength curve, compare strength against embodied-carbon estimates and suggest promising new experiments.[4]
The search can respect local constraints, including limits on the water-to-binder ratio or the absence of a particular ingredient. That matters because a recipe must use materials a supplier can obtain. Each measured result provides another example connecting a composition with its behaviour.[4]
What the open research contains
The March 2026 BOxCrete research dataset
| Measure | Reported value | Meaning |
|---|---|---|
| Mixtures | 123: 69 mortar and 54 concrete | Prepared at the University of Illinois laboratorySource 3 |
| Strength observations | 533 | Measurements across the experimental mixturesSource 3 |
| Curing ages | 1, 3, 5, 14 and 28 days | Tracks early and later strengthSource 3 |
| Held-out evaluation | 12 concrete mixes in each of five random splits | Excluded from training within each splitSource 3 |
| Average prediction error | 0.69 ksi, equivalent to 690 psi | Root-mean-square error across the five test sets and all agesSource 3 |
The preprint’s original optimisation focused on compressive strength and embodied carbon. Workability and durability were outside that model’s scope. A promising numerical prediction therefore still needed practical testing before site use.[3]
The building site adds its own requirements
A data-centre slab must also be flat, smooth and suitable for placing and servicing heavy equipment. Amrize and Mortenson tested whether the new material could be handled and finished properly. Human specialists reviewed the proposed mixes and adjusted constraints during the progression to a usable slab.[1]
That sequence connects material discovery with construction. The model helps the team explore recipes; trial batches and slab tests establish how a selected recipe behaves when people mix, place and finish it.[2]
A reusable tool for local materials
The March 2026 release added more robustness to noisy data and the ability to predict slump, an indicator of fresh concrete’s workability. Meta also reported that concrete-software provider Quadrel had adapted its AI framework for customer-specific training and everyday quality-control workflows.[2]
The current repository includes code, measured mixture data and worked examples under the MIT licence. Its examples show how users can compare a strength curve, explore a carbon-versus-strength trade-off and nominate another laboratory experiment. The reusable result is a way to develop and refine a recipe around the materials and performance requirements of a particular project.[4]
Sources & further reading
- Using AI to make lower-carbon, faster-curing concrete Engineering at Meta · 16 Jul 2025 · Accessed 11 Oct 2026
- AI for American-Produced Cement and Concrete Engineering at Meta · 30 Mar 2026 · Accessed 11 Oct 2026
- BOxCrete: A Bayesian Optimization Open-Source AI Model for Concrete Strength Forecasting and Mix Optimization arXiv / University of Illinois Urbana-Champaign and Meta · 23 Mar 2026 · Accessed 11 Oct 2026
- BOxCrete: Bayesian Optimization for Sustainable Concrete Mix Design Meta / GitHub · Accessed 11 Oct 2026
- Sustainable Concrete via Bayesian Optimization arXiv / Meta and University of Illinois Urbana-Champaign · 20 Nov 2023 · Accessed 11 Oct 2026