Why AI for Concrete Matters Right Now

Concrete is the second most consumed material on Earth after water, and the U.S. pours roughly 400 million cubic yards of it every year. Yet the process of designing a concrete mix remains stubbornly analog: trial-and-error in the lab, engineer intuition, and decades of accumulated know-how. That workflow is slow, expensive, and brittle—especially when you need to swap imported cement for domestically sourced materials.

Meta’s new open-source model, BOxCrete (Bayesian Optimization for Concrete), directly tackles this bottleneck. Instead of brute-force testing, it uses adaptive experimentation to intelligently explore the formulation space. The result? A mix that reached full structural strength 43% faster than the original formula while reducing cracking risk by nearly 10%—and it was used at scale in a real data center foundation in Rosemount, MN.

This isn’t a research toy. It’s a production-grade tool that’s already been embedded into daily workflows by companies like Quadrel, a leading ready-mix SaaS platform. For a deeper look at how privacy-preserving AI techniques can be applied to other infrastructure domains, check out our related piece on how advanced browsing protection checks URLs without compromising privacy.

Meta's Bayesian optimization AI model for concrete mix design on a laptop screen IT Technology Image

How BOxCrete Works: Adaptive Experimentation in Practice

At its core, BOxCrete is a Bayesian optimization engine built on Meta’s Ax platform. Here’s the pipeline:

  1. Learn from existing data – Historical mix designs, lab results, and performance metrics train the model on what works.
  2. Propose high-potential candidates – The AI suggests new mixes most likely to meet target specifications (strength, workability, cost, sustainability).
  3. Incorporate constraints upfront – Users specify technical requirements and which ingredients to use (e.g., only U.S.-produced cement).
  4. Refine with each test – Every lab result improves the model’s predictions, creating an automatic improvement loop.
# Simplified example of BOxCrete’s core logic (conceptual)
from ax import optimize

def concrete_performance(mix_params):
    # Simulate lab test: returns a score based on strength, slump, cost
    # In production, this calls a real model trained on thousands of samples
    strength = predict_28day_strength(mix_params)
    slump = predict_slump(mix_params)  # workability indicator
    cost = estimate_material_cost(mix_params)
    return (strength + slump) / cost  # multi-objective reward

best_params, best_value = optimize(
    parameters=[
        {"name": "cement_type", "type": "choice", "values": ["Type I", "Type III", "Blended"]},
        {"name": "water_cement_ratio", "type": "range", "bounds": [0.35, 0.55]},
        {"name": "fly_ash_pct", "type": "range", "bounds": [0, 30]},
    ],
    evaluation_function=concrete_performance,
    total_trials=50,
)

The key innovation over Meta’s earlier models is robustness to noisy data and the ability to predict concrete slump—a critical indicator of workability that’s notoriously hard to model. The model also comes with a foundational dataset from the Rosemount data center project, which is the best systematic open-source dataset for concrete mix performance currently available.

Data center foundation construction using AI-optimized concrete mix in Minnesota Coding Session Visual

Real-World Impact: From Minnesota to Pennsylvania

Minnesota: Faster, Stronger, Domestic

At Meta’s Rosemount data center, BOxCrete generated a mix that:

  • Reached full structural strength 43% faster than the original formula
  • Reduced cracking risk by nearly 10%
  • Used only domestically sourced materials

The mix was poured in a site support section of the building slab—one of the most demanding parts of the build, supporting thousands of servers and cooling systems. After confirming all structural requirements, the mix is now qualified for additional areas.

Pennsylvania: Embedded in Daily QC Workflows

Quadrel, a Pennsylvania-based SaaS provider for the ready-mix industry, adapted Meta’s AI framework into its platform. Their implementation covers:

  • Data preprocessing and batch normalization
  • Test normalization and feature engineering
  • Customer-specific model training

Field test results continuously improve the models, and they’re now part of daily mix design and quality control decisions.

Illinois: Academic-Industry Partnership

Meta partnered with the University of Illinois at Urbana-Champaign and Amrize (North America’s largest cement manufacturer) to validate the approach at industrial scale. Amrize recently launched a “Made in America” cement label and announced nearly $1B in capital investments to boost domestic production.

Limitations and Caveats

  • Lab validation is still required. BOxCrete accelerates discovery but doesn’t replace physical testing, engineering sign-off, or code compliance.
  • Data quality matters. The model is only as good as the training data. Noisy or incomplete historical data can lead to suboptimal suggestions.
  • Material variability. Different cements have different chemistries; a mix optimized for one supply chain may not transfer directly to another.

What’s Next?

Meta is planning to expand collaboration with the construction industry and academia. Future directions include:

  • Addressing broader sustainability challenges beyond material substitution
  • Reducing embodied carbon further through AI-optimized supplementary materials
  • Making the tool accessible to smaller producers via platforms like Quadrel

For a broader perspective on how AI is moving from prototypes to production in other domains, see our analysis of Vercel’s new v0 platform for AI-powered coding.

Open source AI framework for sustainable concrete production on GitHub Algorithm Concept Visual

Conclusion: AI as a Catalyst for Domestic Supply Chains

BOxCrete isn’t just a technical achievement—it’s a practical tool for reshoring. By reducing the time and cost of reformulating mixes around U.S.-made materials, it helps American producers compete on cost, reduce emissions, and build supply chain resilience. The model is open source (MIT license) and available on GitHub, along with the foundational dataset.

Get started today:

  • Explore the BOxCrete repository on GitHub
  • Read the pre-print: “BOxCrete: A Bayesian Optimization Open-Source AI Model for Concrete Strength Forecasting and Mix Optimization”
  • Join the conversation at the ACI Spring Convention 2026

If you’re a concrete producer or a software vendor serving the construction industry, now is the time to start experimenting. The barrier to entry has never been lower.

This content was drafted using AI tools based on reliable sources, and has been reviewed by our editorial team before publication. It is not intended to replace professional advice.