Problem

Modern concrete mix designs are increasingly complex due to evolving demands for mechanical performance, workability, durability, sustainability, and cost-effectiveness. Traditional methods rely on time-consuming and resource-intensive trial-and-error approaches, which are inefficient for navigating the high-dimensional interdependencies of modern concrete ingredients. While Artificial Intelligence (AI) and Machine Learning (ML) models show promise for predicting compressive strength and guiding mix optimization, most existing efforts are based on proprietary industrial datasets and closed-source implementations. There is a critical need for publicly available, high-fidelity datasets and open-source AI models that can accurately forecast strength development over time and quantify associated uncertainties for robust structural applications.

Method

The authors introduce BOxCrete, an open-source probabilistic modeling and optimization framework designed for concrete strength forecasting and mix optimization. BOxCrete is trained on a new open-access dataset comprising 533 unique strength measurements from 123 experimentally cast mortar (69) and concrete (54) mixes, developed at a single laboratory and tested at five curing ages (1, 3, 5, 14, and 28 days). The AI model leverages Gaussian Process (GP) regression to probabilistically model strength development as a function of time and mix composition, enabling uncertainty-aware predictions. This framework is coupled with Multi-Objective Bayesian Optimization (BO) to perform multi-objective optimization of compressive strength and Global Warming Potential (GWP), aiming to identify optimal binder formulations. The model was iteratively refined across six experimental phases, each incorporating increasingly diverse data, and was rigorously validated against 12 independent concrete mixes excluded from training. Materials used included Portland Limestone Cement, various Class C and F fly ashes, and Grade 100 slag. Compressive strength was tested following ASTM C109 and C39 standards, and GWP was assessed using the OpenConcrete life cycle assessment tool.

Results

BOxCrete demonstrated high predictive accuracy, achieving an average R² of 0.94 and an RMSE of 0.69 ksi for strength prediction across all curing ages. The model successfully reproduced the characteristic sigmoidal strength-evolution behavior in concrete, accurately capturing both rapid early-age strength gains and gradual later-age increases. Predictive accuracy improved systematically with the size and compositional diversity of the training dataset, evidenced by uncertainty bounds contracting from approximately ±2.4 ksi in early phases to ±0.2 ksi by Phase VI. Compared to prior AI-based models, BOxCrete achieved similar or superior accuracy with significantly fewer data points, highlighting its data efficiency and the benefits of its deliberate compositional diversification and phase-wise training strategy. Furthermore, coupling BOxCrete with Bayesian Optimization enabled the identification of optimized concrete mixes achieving compressive strengths exceeding 6 ksi at 28 days while maintaining embodied carbon levels as low as 120–150 kg CO2e/yd³. Specifically, it identified formulations exceeding 10 ksi compressive strength at 28 days with GWP between 150–200 kg CO2e/m³, representing over 50% cement replacement and up to 60% lower GWP than equivalent control mixes. The framework also facilitates inverse design, allowing users to generate candidate mixtures based on predefined strength and GWP constraints.

Implications

BOxCrete establishes a reproducible, open-source foundation for data-driven development of AI-based optimized concrete mix designs, significantly reducing barriers to AI adoption in the construction industry. By providing a publicly available dataset and an open-source model under the MIT license, this work enables accelerated discovery of high-performance, low-carbon concrete formulations and encourages community-driven refinement and broader applicability. The model's probabilistic architecture offers interpretable, uncertainty-aware predictions, empowering decision-makers to quantify confidence levels and efficiently explore strength–sustainability trade-offs. While the current model focuses on specific ingredient types and primarily optimizes mechanical performance and sustainability, future studies are encouraged to expand the dataset to account for the quality and source of various ingredients and incorporate workability and durability as additional optimization parameters. This open-source approach paves the way for integrating AI into commercial software to advance sustainable concrete mix design globally.