Explainable Artificial Intelligence for Concrete Mix Design Using Physics-Grounded Synthetic Data

Abstract

Machine learning has shown strong potential for predicting concrete compressive strength based on mixture composition, yet its adoption in engineering practice is limited by poor interpretability. This study investigates the use of explainable artificial intelligence to understand and validate machine learning decisions in concrete mix design. A physics consistent synthetic dataset is generated to represent realistic concrete mixtures, ensuring that governing relationships such as the influence of cement content and water cement ratio are known a priori. A gradient boosting regression model is trained to predict compressive strength and subsequently interpreted using multiple explainability techniques, including global feature attribution, local explanation analysis, response curves, and interaction visualizations. The results demonstrate that the model successfully recovers established concrete mechanics principles, including the dominant role of cement content, the inverse effect of water cement ratio, and the secondary influence of curing age, temperature, and admixture dosage. By validating explanations against known synthetic ground truth, this work shows that explainable artificial intelligence can serve as both an interpretation and validation framework, enabling transparent and physically meaningful data driven concrete mix design.

Introduction

Traditional concrete mix design methods are empirical, time-consuming, and unable to fully capture complex nonlinear relationships between ingredients and strength. Machine learning improves prediction by modeling these nonlinear relationships using variables such as cement content, water content, curing age, and temperature. However, most ML models are “black boxes,” making it difficult to understand how predictions are made, which limits their use in engineering practice.

To address this, the study applies explainable AI techniques to a physics-consistent synthetic concrete dataset where true relationships are known. A Gradient Boosting Regressor is trained to predict compressive strength, and multiple XAI methods (such as SHAP, partial dependence plots, and feature interaction analysis) are used to interpret the model.

Results show that the model correctly identifies key physical relationships:

• Cement content has the strongest positive effect on strength
• Water–cement ratio has a strong negative effect (confirming Abrams’ law)
• Curing age and admixture dosage contribute positively but moderately
• Aggregate and water content have minimal direct impact

Conclusion

This study demonstrated the effectiveness of explainable artificial intelligence (XAI) for interpreting and validating machine learning–based concrete mix design models using a physics-consistent synthetic dataset. Rather than focusing solely on predictive accuracy, the work emphasized understanding why specific mixture compositions lead to higher or lower compressive strength, addressing a key limitation that often restricts the practical adoption of machine learning in civil engineering. Global explainability analysis using SHAP revealed that cement content and water–cement ratio are the dominant variables governing compressive strength prediction. The opposing SHAP trends of these two parameters confirmed that the trained model independently rediscovered fundamental concrete mechanics, particularly Abrams’ law, without being explicitly constrained to do so. Secondary variables such as curing age, admixture dosage, and temperature exhibited moderate but physically meaningful contributions, while aggregate and absolute water content showed minimal direct influence, consistent with their known roles in concrete behavior. Local SHAP explanations further demonstrated that high-strength concrete mixtures are achieved through a coherent combination of high cement content and low water–cement ratio, supported by favorable curing age and temperature conditions. The clear hierarchy of feature contributions in the local waterfall analysis confirmed that the model’s decision-making process is transparent and aligns with established engineering intuition at the individual mix level. Partial dependence and accumulated local effects analyses provided complementary global insights, highlighting monotonic and stable response trends across key variables. Cement content showed strong positive influence with diminishing marginal gains at higher dosages, while water–cement ratio exhibited a consistently negative effect on strength. Admixture dosage contributed positively but modestly, reinforcing its secondary role in strength development. Two-dimensional PDP and SHAP interaction analyses showed that interaction effects between cement and water are generally weak and localized, indicating that compressive strength is largely governed by additive effects and relative proportions rather than strong nonlinear interactions between absolute quantities.

References

[1] Shi, C., Wu, Z., Lv, K. and Wu, L., 2015. A review on mixture design methods for self-compacting concrete. Construction and Building Materials, 84, pp.387-398. [2] Ahmad, S. and Alghamdi, S.A., 2014. A statistical approach to optimizing concrete mixture design. The scientific world journal, 2014(1), p.561539. [3] DeRousseau, M.A., Kasprzyk, J.R. and Srubar Iii, W.V., 2018. Computational design optimization of concrete mixtures: A review. Cement and concrete research, 109, pp.42-53. [4] Li, N., Shi, C., Zhang, Z., Wang, H. and Liu, Y., 2019. A review on mixture design methods for geopolymer concrete. Composites Part B: Engineering, 178, p.107490. [5] Kosmatka, S.H., Panarese, W.C. and Kerkhoff, B., 2002. Design and control of concrete mixtures (Vol. 5420, pp. 60077-1083). Skokie, IL: Portland cement association. [6] Yeh, I.C., 2007. Computer-aided design for optimum concrete mixtures. Cement and Concrete Composites, 29(3), pp.193-202. [7] Jiao, D., Shi, C., Yuan, Q., An, X. and Liu, Y., 2018. Mixture design of concrete using simplex centroid design method. Cement and Concrete Composites, 89, pp.76-88. [8] Shi, C., Wu, Z., Xiao, J., Wang, D., Huang, Z. and Fang, Z., 2015. A review on ultra high performance concrete: Part I. Raw materials and mixture design. Construction and Building Materials, 101, pp.741-751. [9] Hüsken, G. and Brouwers, H.J.H., 2008. A new mix design concept for earth-moist concrete: A theoretical and experimental study. Cement and Concrete Research, 38(10), pp.1246-1259. [10] Tahwia, A.M., Hamido, M.A. and Elemam, W.E., 2023. Using mixture design method for developing and optimizing eco-friendly ultra-high performance concrete characteristics. Case Studies in Construction Materials, 18, p.e01807. [11] Sun, C., Wang, K., Liu, Q., Wang, P. and Pan, F., 2023. Machine-learning-based comprehensive properties prediction and mixture design optimization of ultra-high-performance concrete. Sustainability, 15(21), p.15338. [12] Ziolkowski, P. and Niedostatkiewicz, M., 2019. Machine learning techniques in concrete mix design. Materials, 12(8), p.1256. [13] Zheng, W., Shui, Z., Xu, Z., Gao, X. and Zhang, S., 2023. Multi-objective optimization of concrete mix design based on machine learning. Journal of Building Engineering, 76, p.107396. [14] Alghamdi, S.J., 2023. Determining the mix design method for normal strength concrete using machine learning. Journal of Umm Al-Qura University for Engineering and Architecture, 14(2), pp.95-104. [15] Hosseinnia, A., Sichani, M.N., Alamdari, B.E., Aghelizadeh, P. and Teimortashlu, A., 2025, June. Machine learning formulation for predicting concrete carbonation depth: A sustainability analysis and optimal mixture design. In Structures (Vol. 76, p. 109036). Elsevier.

Copyright

Copyright © 2026 Rahul Gocher, Arun Kumar Yadav. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.