BOxCrete: A Bayesian Optimization Open-Source AI Model for Concrete Strength Forecasting and Mix Optimization

Bayezid Baten; M. Ayyan Iqbal; Sebastian Ament; Julius Kusuma; Nishant Garg

cs.LGMar 23, 2026

BOxCrete: A Bayesian Optimization Open-Source AI Model for Concrete Strength Forecasting and Mix Optimization

Authors: Bayezid Baten, M. Ayyan Iqbal, Sebastian Ament, Julius Kusuma, Nishant Garg

Organizations: Department of Civil and Environmental Engineering, University of Illinois Urbana-Champaign, Urbana, IL, USA · Meta, Menlo Park, California, USA.

Abstract

Modern concrete must simultaneously satisfy evolving demands for mechanical performance, workability, durability, and sustainability, making mix designs increasingly complex. Recent studies leveraging Artificial Intelligence (AI) and Machine Learning (ML) models show promise for predicting compressive strength and guiding mix optimization, but most existing efforts are based on proprietary industrial datasets and closed-source implementations. Here we introduce BOxCrete, an open-source probabilistic modeling and optimization framework trained on a new open-access dataset of over 500 strength measurements (1-15 ksi) from 123 mixtures - 69 mortar and 54 concrete mixes tested at five curing ages (1, 3, 5, 14, and 28 days). BOxCrete leverages Gaussian Process (GP) regression to predict strength development, achieving average R $^2$ = 0.94 and RMSE = 0.69 ksi, quantify uncertainty, and carry out multi-objective optimization of compressive strength and embodied carbon. The dataset and model establish a reproducible open-source foundation for data-driven development of AI-based optimized mix designs.

Figures & tables

Figure 1: AI-in-the-loop framework of BOxCrete for sustainable concrete mix optimization. Raw materials (Type IL cement, SCMs, and aggregates) are proportioned into mortar and concrete mixes, cast, and tested for compressive strength at 1, 3, 5, 14, and 28 days. Cradle-to-gate Global Warming Potential (GWP) is quantified via life cycle assessment (LCA). The resulting dataset of mix proportions, strength, and GWP is used to train a Gaussian Process Regression model to forecast strength evolution with quantified uncertainty. Coupled with Bayesian Optimization (BoTorch), the trained model generates new mix designs targeting user-defined strength and sustainability objectives. AI-suggested formulations are experimentally validated and reintroduced into the loop, enabling adaptive learning and accelerated discovery of high-performance, sustainable concrete for infrastructure applications.

Figure 2: Evolution of compressive strength and embodied carbon in mortar and concrete mixes across curing ages. Global warming potential (GWP; kg CO 2 e/yd 3 ) is plotted against unconfined compressive strength ( $f^{\prime}_{c}$ ; ksi) for 69 mortar (circles) and 54 concrete (triangles) mixes at 1, 3, 5, 14, and 28 days of curing. Each data point represents a unique mix design with varying binder content (240–2150 lb/yd 3 ), water-to-binder (w/b) ratios (0.20–0.50), and different replacement levels of Portland Limestone Cement with Class C and Class F fly ash and ground granulated blast furnace slag. Vertical error bars indicate the standard error of the compressive strength measurements based on three replicate specimens. GWP values were computed on a cradle-to-gate basis using standardized material emission factors. The optimized performance zone, defined by low embodied carbon and high compressive strength, corresponds to the top-right quadrant of each plot.

Figure 3: Evolution of AI model predictions for compressive strength as a function of training dataset size. Predicted (blue lines) and experimentally measured (green points) compressive strength ( $f^{\prime}_{c}$ , ksi) are shown for six representative concrete mixes across six training dataset phases (Phase I to Phase VI). Each phase corresponds to an expanded training set size, incorporating an increasing number of mix designs with varying binder compositions and water-to-binder (w/b) ratios. Shaded regions indicate the 95% confidence interval (mean $\pm 2$ standard deviations) of model predictions. Experimental data points represent the mean $\pm 1$ standard deviation from triplicate tests ( $n=3$ ). The figure illustrates progressive improvement in model accuracy and uncertainty reduction as the training set becomes more comprehensive.

Figure 4: Phase-wise improvement in BOxCrete predictive accuracy for Testing Set 1. Model performance is shown across six training phases (P-I to P-VI) using compressive-strength data from 123 mortar and concrete mixes evaluated at 1, 3, 5, 14, and 28 days. Left: Final-phase (P-VI) parity plot comparing predicted ( $f^{\prime}_{c,p}$ ) and experimental ( $f^{\prime}_{c,e}$ ) strengths for $n=60$ independent test samples ( $R^{2}=0.94$ , RMSE $=0.73$ ksi). Marker shapes denote curing age; error bars represent $\pm 1$ standard deviation (total span $=2\sigma$ ). Strong alignment with the 1:1 line indicates high predictive accuracy across hydration stages. Right: Phase-specific parity plots (P-I to P-VI) and aggregated panel (P-I to VI) demonstrate progressive reduction in scatter and bias, with early phases underestimating high-strength mixes and later phases converging closely to parity. Bottom: Evolution of $R^{2}$ and RMSE with phase iteration for each curing age shows consistent improvement, with the largest gains at 14 and 28 days due to improved representation of SCM-driven delayed hydration reactions.

Metric	Age	Set 1	Set 2	Set 3	Set 4	Set 5	All Sets
R 2	1-Day	0.67	0.94	0.86	0.86	0.67	0.80 $\pm$ 0.12
	3-Days	0.69	0.82	0.82	0.86	0.69	0.78 $\pm$ 0.08
	5-Days	0.83	0.83	0.86	0.82	0.83	0.83 $\pm$ 0.02
	14-Days	0.92	0.85	0.87	0.93	0.92	0.90 $\pm$ 0.04
	28-Days	0.91	0.84	0.84	0.93	0.91	0.89 $\pm$ 0.04
	All Ages	0.94	0.94	0.93	0.94	0.96	0.94 $\pm$ 0.01

Table 1: Performance metrics of the BOxCrete model across testing datasets and curing ages. Statistical performance across five independent testing sets is shown in terms of the Coefficient of Determination (R 2 ) and Root-Mean-Square Error (RMSE) for 1-, 3-, 5-, 14-, and 28-day curing ages. The global averages yield R 2 = 0.94 ± 0.01 and RMSE = 0.69 ± 0.07 ksi.

Figure 5: Compositional diversity and binder synergy governing strength development. Scatter plots show the distribution of all mortar and concrete mixtures, relating 28-day compressive strength ( $f^{\prime}_{c}$ ) to key mix parameters including cement, fly ash, slag, aggregates, HRWR dosage, paste content, SCM replacement, and water-to-binder (w/b) ratio. The broad, non-monotonic trends reflect the intentionally diverse design space spanning both high- and low-clinker systems, forming the basis for robust model training. Ternary diagrams depict binder phase distributions (cement–slag–fly ash) at 1, 5, and 28 days, with strength indicated by a continuous color gradient. Early-age strength concentrates in cement-rich compositions, whereas later-age high-strength regions shift toward balanced ternary blends, highlighting progressive SCM activation and synergistic hydration effects.

Phase	Mixes	Type	Parameters Modified
P I	69 (M1–M69)	Mortar	Binder content, w/b ratio, cement replacement, HRWR dosage, curing temp.
P II	7 (C1–C4, C6–C8)	Concrete	Fine-to-coarse ratio
P III	5 (C10–C14)	Concrete	SCM dosage, w/b ratio, HRWR content
P IV	10 (C16–C20, C22–C25)	Concrete	Binder content, w/b ratio, SCM dosage, HRWR content
P V	4 (C28–C31)	Concrete	Binder content, cement replacement, HRWR content
P VI	16 (C33–C53)	Concrete	Binder content, cement replacement, w/b ratio, HRWR dosage, curing temp.

Table 2: Summary of Experimental Mix Development Phases and Key Variable Adjustments for Training Set 1.

Materials Property Prediction

Materials property prediction uses data and physical models to estimate how composition, structure, and processing determine a material’s behavior.

See all 165 papers

Reactivity-Informed Machine Learning for Performance Prediction and Design Space Exploration of Alkali-Activated Slag
Jun 4, 2026Qiyao He, Zhanzhao Li, Kai GongMaterials Property Prediction Materials Science
Department of Civil and Environmental Engineering, Rice University, Houston, Texas 77005, United States. · Rice Advanced Materials Institute, Rice University, Houston, Texas 77005, United States · Ken Kennedy Institute, Rice University, Houston, Texas 77005, United States
Large language model-enabled automated data extraction for concrete materials informatics
Apr 24, 2026Zhanzhao Li, Kengran Yang, Qiyao He +1LLM Information Extraction
Department of Civil and Environmental Engineering, Rice University, Houston, TX 77005, USA. · Rice Advanced Materials Institute, Rice University, Houston, TX 77005, USA. · Ken Kennedy Institute, Rice University, Houston, TX 77005, USA. +1
Incremental Transformer for Surrogate-Based Inverse Design of Geopolymer Mixtures
Jul 12, 2026Giansalvo Cirrincione, Filippo GrassiaSurrogate Modeling Materials Property Prediction
Laboratory of Innovative Technologies (LTI, EA 3899) University of Picardie Jules Verne, Amiens, 80025, Hauts-de-France, France

Jun 4, 2026J/K move · Enter open · S save

Reactivity-Informed Machine Learning for Performance Prediction and Design Space Exploration of Alkali-Activated Slag

Qiyao He, Zhanzhao Li, Kai Gong

Department of Civil and Environmental Engineering, Rice University, Houston, Texas 77005, United States. · Rice Advanced Materials Institute, Rice University, Houston, Texas 77005, United States · Ken Kennedy Institute, Rice University, Houston, Texas 77005, United States

Materials Property Prediction Materials Science

Open paper AskPDF arXiv

Establishing quantitative relationships among mix design, raw material properties, curing conditions, and performance remains a long-standing challenge in cementitious materials, particularly for alkali-activated materials with variable precursor and activator chemistry. Here, we curated the largest literature-derived alkali-activated slag (AAS) dataset to date, comprising over 3100 compressive strength records, 155 chemically distinct ground granulated blast-furnace slags (GGBSs), and 24 attributes incorporating precursor chemistry, fineness, and reactivity. Multiple machine learning (ML) algorithms were benchmarked across progressively enriched feature scenarios, demonstrating that integrating GGBS compositions, fineness, curing conditions, and specimen geometry improves predictive performance. The average metal oxide dissociation energy (AMODE), a physically interpretable representation of precursor reactivity, provides a compact alternative descriptor to explicit oxide compositions while enabling comparable predictive performance. Model interpretation revealed physically consistent trends from heterogeneous data, including non-monotonic effects of Na2O dosage and silicate modulus, reduced predicted strength at higher water content and larger specimen size, and coupled oxide-level effects more coherently represented by AMODE than by individual oxide contents. Statistically constrained design space exploration reveals reactivity-dependent trade-offs among strength, embodied CO2 emissions, and cost. The design maps identify high-strength regions with substantially lower CO2 emissions than OPC-based references at similar cost. Overall, this work demonstrates how reactivity-informed ML can extract physically meaningful trends from heterogeneous AAS data and guide source-dependent binder design. The curated dataset is publicly accessible to support advances in cement and concrete research.

Abstract

Materials Property Prediction

Explore similar work

Reactivity-Informed Machine Learning for Performance Prediction and Design Space Exploration of Alkali-Activated Slag

Large language model-enabled automated data extraction for concrete materials informatics

Incremental Transformer for Surrogate-Based Inverse Design of Geopolymer Mixtures