Design of concrete mixtures and prediction of their compressive strength using machine learning

Design of concrete mixtures and prediction of their compressive strength using machine learning Radoslav Gandel1 * , Jan Jerabek1, Petr Cmiel2, and Oldrich Sucharda1 1VSB – Technical University of Ostrava, Faculty of Civil Engineering, Department of Building Materials and Diagnostics of Structures, Ludvika Podeste 1875/17, 708 00 Ostrava-Poruba, Czech Republic 2TESTSTAV, spol. s r.o., Františka Lyska 1599/6, 700 30 Ostrava – Belsky Les,Czech Republic Abstract. The use of machine learning and neural networks in predicting the compressive strength of concrete promises to significantly improve the accuracy and reliability of models for the design and optimization of concrete mixtures. With rapid advances in this field, computational models will be able to handle even larger amounts of experimental data, increasing their ability to capture the complex relationships between input parameters and the mechanical properties of concrete. With the development of new neural network architectures and machine learning algorithms, it will be possible to create highly adaptive predictive models that can better respond to variability in concrete composition and production conditions, leading to more efficient and sustainable design in the construction industry. The submitted paper deals with the design of concrete mixtures and prediction of their compressive strength based on the compressive strength results of mixtures of known composition from other experiments using machine learning. Practical validation of the developed regression model will be carried out by testing the machine-designed mixtures for compressive strength after 28 days. 1 Introduction Concrete is one of the most widely used building materials in the world due to its availability, properties and wide range of applications [1, 2]. Despite its widespread use, the proper design of concrete mixture is a challenge because its properties depend not only on the nature of the input materials and the production technology, but also on the maturation time of the mixture. In the construction industry, accurate prediction of the key properties of concrete is important not only for the safety and reliability of structures [3], but also for reducing costs and carbon footprint [4, 5]. The methods used to predict strength characteristics, such as experimental tests, are time and cost consuming. Therefore, many researches are currently exploring the use of modern, advanced methods such as machine learning [6-8] or deep learning using neural networks [9, 10] to predict concrete properties. Machine learning, as one of the most * Corresponding author: ra[email protected] E3S Web of Conferences 641, 01026 (2025) https://doi.org/10.1051/e3sconf/202564101026 Young Scientist 2025 © The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0 (https://creativecommons.org/licenses/by/4.0/). fundamental tools in the world of artificial intelligence, represents an innovative and relatively simple approach to solving civil engineering problems. Its main advantage, based on the analysis and processing of large amounts of data, is the ability to identify relationships and patterns obscure to conventional methods. In the practice of concrete mixture design, this means eliminating experimental testing while reducing financial costs and time, which is achieved by predicting the desired properties based on the composition of the concrete and its eventual subsequent optimization. Although machine learning is easier to understand compared to more advanced artificial intelligence methods, its effectiveness is largely influenced by several negative factors: the quality of the dataset, the time and cost associated with creating the model and the limited ability to handle complex relationships. Nevertheless, it has several advantages over deep learning, such as lower computational requirements or the suitability of using even smaller datasets. The issue of predicting the properties of concrete is particularly topical given the increasing demands for sustainability in the construction industry [11, 12]. The use of artificial intelligence in construction is a modern approach that can meet the requirements of reducing CO₂ emissions, efficient use of resources and improving concrete properties. After understanding the relationships between the constituents of simpler concrete mixtures, it would be possible to predict the properties of even more complex mixtures for computational models [13-18]. The aim of this paper is to validate the accuracy of predicting the compressive strength of concrete mixtures proposed by the developed machine learning model based on compressive strength data from other researches [19]. 2 Experimental program The making of the computational model was carried out in the Python programmatic language, using as a training dataset [19] with more than 1000 values of the compressive strength of cylinders with dimensions ⌀150x300 mm calculated for different types of concrete mixtures with variable composition consisting of the following input materials: ordinary portland cement type I (classification according to ASTM C150), blast furnace slag, fly ash, water, coarse aggregate (up to 10 mm), fine aggregate (fineness module of 3) and plasticiser based on naphthalene-formaldehyde condensate and fatty acid copolymer. Cylinder compressive strength was subsequently measured at a time horizon of 1 - 365 days of age of the samples. After programming the computational model, a user-friendly environment was created, an example of which is shown in Fig. 1. Fig. 1. Environment of the app for predicting the cylinder compressive strength of concrete. E3S Web of Conferences 641, 01026 (2025) https://doi.org/10.1051/e3sconf/202564101026 Young Scientist 2025 2 The user interface of the app (shown in Fig. 1) consists of a toolbar where the weight range of independent variables (input materials of the concrete mixture) and the age of the samples can be set. On the right side is displayed the volume of the concrete mix calculated based on the amount of input materials entered by the user. The individual components per 1 m3 of fresh concrete mix and the predicted cylinder compressive strength for the specified sample age are then calculated, along with a graphical, predicted development of cylinder compressive strength over time. 2.1 Programming of the computational model After training and testing several types of regression models, the Light Gradient Boosting Machine model with root mean square error RMSE = 4.9378 (MPa) and coefficient of determination R2 = 0.9046 was chosen as the most appropriate model. After fine tuning of the tested models with 10 iterations and 100 folds and final fine tuning of the best fitting regression models with 5 iterations and 5 folds, the Gradient Boosting Regressor model with RMSE = 3.9406 (MPa) and R2 = 0.9491 was selected as the optimal model, indicating the high ability of the model to explain the variability in the data with a relatively small deviation of the predicted cylinder compressive strength from the actual cylinder compressive strength. The quality of the prediction of the regression model could also be explained by the graph of prediction errors in Fig. 2. Fig. 2. Graph of prediction errors of Gradient Boosting Regressor. The graph in Fig. 2 shows two lines: the line of best fit and the line of identity. The line of identity represents the line of perfect prediction, where the predicted cylinder compressive strength values are equal to the actual ones. The line of best fit represents the best possible linear approximation that tries to best describe the relationship between the predicted cylinder compressive strength values and the actual ones. Most of the points in the graph are close to the identity line and the line of best fit has a small deviation from it, indicating the high quality of the model. E3S Web of Conferences 641, 01026 (2025) https://doi.org/10.1051/e3sconf/202564101026 Young Scientist 2025 3 2.2 Verification of the computational model on real samples As mentioned in the previous chapter, the user environment contains a panel with input materials, whose interval represents the minimum and maximum values of a given input material obtained from the corresponding dataset [19]. After the user enters the weights of individual input materials, the given weights per 1 m3 of concrete mix are proportionally recalculated according to the respective volumetric weights of the materials. Since the oneday compressive strengths in the dataset contain only two values, the panel for the age of the sample is set from third day. For the experimental testing of the computational model, two mixtures were randomly designed without any input requirements (the composition of which is given in Table 1). For each mixture, 7 cube samples with an edge length of 150 mm were prepared for logistical reasons, on which the cube compressive strength was performed after 28 days (due to standard CSN EN 12390-3 [20]) and subsequently converted to the cylinder compressive strength (empirically determined as 80-85 % of the cube compressive strength). 28-day cube samples of both mixtures are illustrated in Fig. 3. In the practical part of the experiment, the used cement was portland cement CEM I 42.5 Hranice, in the case of blast furnace slag - granulated blast furnace slag JMS 420 Stramberk, then coarse aggregate 4-8 mm Litice, fine aggregate 0-4 mm Tovacov and a plasticizer based on polycarboxylate ethers. Table 1. Designed concrete mixtures. Input materials Recalculated quantity kg∙m-3 ML1 ML4 Cement 377.44 571.59 Blast furnace slag 91.04 - Fly ash - - Water 166.73 187.35 Plasticizer - 17.99 Coarse aggregate 933.88 903.96 Fine aggregate 901.15 747.30 Fig. 3. 28-day samples of mixtures ML1 (left) and ML4 (right). E3S Web of Conferences 641, 01026 (2025) https://doi.org/10.1051/e3sconf/202564101026 Young Scientist 2025 4 Table 2 shows the cylinder compressive strength results for the interval 80-85 % of the cube compressive strength, the predicted compressive strength, the average deviation of the predicted values of the regression model from the actual values, and the average absolute deviation. In the case of the ML1 mixture, it can be observed that the actual absolute deviation (1.49 MPa) is lower than that of the regression model (3.94 MPa), so that in this case the computational model predicted the compressive strength of the cylinder quite accurately, while in the case of the ML4 mixture the absolute deviation (16.03 MPa) is approximately 4 times higher compared to that of the regression model (3.94), indicating its high inaccuracy in predicting the property. Table 2. Observed properties and their statistical characteristics. Mixture Actual cylinder compressive strength [MPa] Predicted cylinder compressive strength [MPa] RMSE [MPa] Average absolute deviation [MPa] ML1 47.72 - 50.70 48.38 3.94 1.49 ML4 43.74 - 46.47 61.13 16.03 3 Conclusion The proposed computational model represents a simplified form of predicting the compressive strength of concrete, dependent on its individual components. The model does not account for additional input parameters and factors influencing the final compressive strength, such as the water-cement ratio, chemical composition, specific surface area and reactivity of binder, mechanical properties of the aggregate, etc. Based on the provided data from the dataset [19], the importance of each variable in the computational model was determined, as its impact on the prediction of compressive strength (see graph in Fig. 4). Fig. 4. The importance of individual variables in the prediction of compressive strength. As can be seen from the graph in Figure 4, the variable cement has the most significant impact on the calculation (excluding the age of the samples), with a relative influence of more than 35%. In contrast, fly ash has a relative influence of approximately 1% on the prediction of compressive strength, which could indicate that it was used in only a small number of E3S Web of Conferences 641, 01026 (2025) https://doi.org/10.1051/e3sconf/202564101026 Young Scientist 2025 5 cases (specifically 464 cases). However, the relative influence of coarse aggregate, which was used in all cases, is around 2%. Thus, these materials, within the specified range of usage quantities, have only a minimal effect on the prediction of compressive strength. To increase the predicted compressive strength, it is essential to prioritize the composition of cement and, to a considerable extent, also focus on water, plasticizers, and blast furnace slag. The final regression model, Gradient Boosting Regressor, with a coefficient of determination R² = 0.9491 and standard deviation RMSE = 3.9406 of the predicted versus actual compressive strength, appears to be suitable. This was confirmed on the ML1 mixture, where the model predicted the compressive strength after 28 days with an average absolute deviation of approximately 1.5 MPa. However, for the ML4 mixture, with an average absolute deviation of approximately 16 MPa between the actual and predicted values, this statement is false. It is also important to note that, in the practical part, the exact same input materials as those used in the dataset were not applied, which significantly affects the compressive strength results. In future work, the model could be expanded by incorporating additional parameters such as water-cement ratio, material-specific properties, or curing conditions. Further development may also include testing more advanced machine learning algorithms and validating the model on larger and more diverse datasets to improve its generalization and practical applicability. The data presented in this study are available on [21]. This paper was supported project No. CZ.02.01.01/00/22_008/0004631 Materials and technologies for sustainable development within the Jan Amos Komensky Operational Program financed by the European Union and from the state budget of the Czech Republic. References 1. P.-C. Aı tcin, Cements of yesterday and today: Concrete of tomorrow, Cement and Concrete Research, 30, 1349-1359 (2000), https://doi.org/10.1016/S00088846(00)00365-3 2. M.C.G. Juenger, F. Winnefeld, J.L. Provis, J.H. Ideker, Advances in alternative cementitious binders, Cement and Concrete Research, 41, 1232-1243 (2011), https://doi.org/10.1016/j.cemconres.2010.11.012 3. S. Zhang, J. Teizer, J.-K. Lee, Ch. M. Eastman, M. Venugopal, Building Information Modeling (BIM) and Safety: Automatic Safety Checking of Construction Models and Schedules, Automation in Construction, 29, 183-195 (2013), https://doi.org/10.1016/j.autcon.2012.05.006 4. B. C. McLellan, R. P. Williams, J. Lay, A. van Riessen, G. D. Corder, Costs and carbon emissions for geopolymer pastes in comparison to ordinary portland cement, Journal of Cleaner Production, 19, 1080-1090 (2011), https://doi.org/10.1016/j.jclepro.2011.02.010 5. C. Meyer, The greening of the concrete industry, Cement and Concrete Composites, 31, 601-605 (2009), https://doi.org/10.1016/j.cemconcomp.2008.12.010 6. D.-Ch. Feng, Z.-T. Liu, X.-D. Wang, Y. Chen, J.-Q. Chang, et al., Machine learningbased compressive strength prediction for concrete: An adaptive boosting approach, Construction and Building Materials, 230, 117000 (2020), https://doi.org/10.1016/j.conbuildmat.2019.117000 E3S Web of Conferences 641, 01026 (2025) https://doi.org/10.1051/e3sconf/202564101026 Young Scientist 2025 6 7. W. B. Chaabene, M. Flah, M. L. Nehdi, Machine learning prediction of mechanical properties of concrete: Critical review, Construction and Building Materials, 260, 119889 (2020), https://doi.org/10.1016/j.conbuildmat.2020.119889 8. H.-T. Thai, Machine learning for structural engineering: A state-of-the-art review, Structures, 38, 448-491 (2022), https://doi.org/10.1016/j.istruc.2022.02.003 9. I.-C. Yeh, Modeling of strength of high-performance concrete using artificial neural networks, Cement and Concrete Research, 28, Issue 12, 1797-1808 (1998), https://doi.org/10.1016/S0008-8846(98)00165-3 10. S.-Ch. Lee, Prediction of concrete strength using artificial neural networks, Engineering Structures, 25, 849-857 (2003), https://doi.org/10.1016/S01410296(03)00004-X 11. H. Naseri, H. Jahanbakhsh, P. Hosseini, F. M. Nejad, Designing sustainable concrete mixture by developing a new machine learning technique, Journal of Cleaner Production, 258, 120578 (2020), https://doi.org/10.1016/j.jclepro.2020.120578 12. M. Shahrokhishahraki, M. Malekpour, S. Mirvalad, G. Faraone, Machine learning predictions for optimal cement content in sustainable concrete constructions, Journal of Building Engineering, 82, 108160 (2024), https://doi.org/10.1016/j.jobe.2023.108160 13. R. Gandel, J. Jerabek, Z. Marcalikova, Reinforced Concrete Beams Without Shear Reinforcement Using Fiber Reinforced Concrete and Alkali-Activated Material, Civil and Environmental Engineering, 19, 348-356, (2023), https://doi.org/10.2478/cee2023-0031 14. O. Sucharda, Z. Marcalikova, R. Gandel, Microstructure, Shrinkage, and Mechanical Properties of Concrete with Fibers and Experiments of Reinforced Concrete Beams without Shear Reinforcement, Materials, 15, 5707, (2022), https://doi.org/10.3390/ma15165707 15. P. Mateckova, V. Bilek, O. Sucharda, Comparative Study of High-Performance Concrete Characteristics and Loading Test of Pretensioned Experimental Beams, Crystals, 11, 427, (2021), https://doi.org/10.3390/cryst11040427 16. Z. Marcalikova, J. Jerabek, R. Gandel, R. Gabor, V. Bilek, O. Sucharda, Mechanical Properties, Workability, and Experiments of Reinforced Composite Beams with Alternative Binder and Aggregate, Buildings, 14, 2142, (2024), https://doi.org/10.3390/buildings14072142 17. J. Nemec, R. Gandel, J. Jerabek, O. Sucharda, V. Bilek, Properties of Selected AlkaliActivated Materials for Sustainable Development, Civil and Environmental Engineering, 20, 307–318, (2024), https://doi.org/10.2478/cee-2024-0024 18. O. Sucharda, R. Gandel, P. Cmiel, J. Jerabek, V. Bilek, Utilization of HighPerformance Concrete Mixtures for Advanced Manufacturing Technologies, Buildings, 14, 2269, (2024), https://doi.org/10.3390/buildings14082269\ 19. I. Yeh, Concrete Compressive Strength [Dataset], UCI Machine Learning Repository (1998), https://doi.org/10.24432/C5PK67 20. ČSN EN 12390-3. Testing hardened concrete – Part 3: Compressive strength of test specimens. Prague: The Czech Office for Standards, Metrology and Testing, 2020 21. R. Gandel, J. Jerabek, P. Cmiel, O. Sucharda, Design of concrete mixtures and prediction of their compressive strength using machine learning (experimental dataset), https://doi.org/10.5281/zenodo.15406920 E3S Web of Conferences 641, 01026 (2025) https://doi.org/10.1051/e3sconf/202564101026 Young Scientist 2025 7