AI for American-Produced Cement and Concrete

Abstraction: Meta's BOxCrete Bayesian optimization model for domestic concrete mix design

Key points:

• BOxCrete (Bayesian Optimization for Concrete) uses Meta's Adaptive Experimentation (Ax) platform to navigate the vast space of concrete formulations; released open-source under MIT license on GitHub
• US imports roughly 20-25% of its cement; BOxCrete helps producers rapidly reformulate around domestically sourced US materials without months of lab trials
• At Meta's Rosemount, MN data center: AI-optimized mix reached full structural strength 43% faster than original formula and reduced cracking risk by nearly 10%
• BOxCrete improvements over prior models: more robustness to noisy data and new ability to predict concrete slump (a workability indicator)
• Pennsylvania-based Quadrel (SaaS platform for ready-mix industry) has embedded Meta's AI framework into daily mix design and quality control workflows
• DS-STAR comparison: with Gemini-2.5-Pro better on hard tasks, GPT-5 better on easy tasks (generalizability finding from DS-STAR, not BOxCrete — but same research context)

Connections: Meta · Amrize · Bayesian Optimization · AI For Science

Source: https://engineering.fb.com/2026/03/30/data-center-engineering/ai-for-american-produced-cement-and-concrete/