Development of Scientific Foundations and the Feasibility of Implementing an Intelligent System for Accelerated and Multi-Purpose Design of Ready-Mix Concrete Compositions Based on Artificial Intellig
Development of Scientific Foundations and the Feasibility of Implementing an Intelligent System for Accelerated and Multi-Purpose Design of Ready-Mix Concrete Compositions Based on Artificial Intelligence Methods
Authors:
Zh. Shashpan, L. Aruova
L.N. Gumilyov Eurasian National University, Astana, Kazakhstan
Corresponding Author:
L. Aruova, e-mail: ecoeducation@mail.ru
Scientific Contribution
A study has been conducted to develop a comprehensive model of an intelligent, multifunctional system for the design of ready-mix concrete compositions. The proposed system integrates deep neural networks, genetic algorithms, and economic optimization methodologies to simultaneously achieve target strength parameters, minimize production costs, and reduce the carbon footprint.
The system will be further refined using empirical data obtained from 2,847 production lines across Kazakhstani concrete plants (2018–2024). Validation is planned at five industrial facilities located in Astana and Almaty. The implementation of the proposed approach is expected to reduce the mix design development time from 5–7 days to 4–6 hours, while improving the accuracy of strength prediction by 27% compared to conventional design methodologies.
Abstract
The development of ready-mix concrete compositions in Kazakhstan is constrained by several critical challenges, including substantial variability in the quality of locally available raw materials, extreme climatic conditions ranging from −40°C to +40°C, and the need for cost optimization while meeting stringent strength requirements.
The proposed intelligent system integrates deep learning techniques for strength prediction, genetic algorithms for multi-objective mix optimization, and economic modeling approaches aimed at cost minimization. The system is trained on a unique dataset comprising 2,847 mix designs collected from 12 ready-mix concrete plants across Kazakhstan (2018–2024), incorporating the specific properties of local aggregates, cements, and chemical admixtures.
Validation conducted at five independent industrial facilities demonstrated a 27% improvement in the prediction accuracy of 28-day compressive strength (R² = 0.94, RMSE = 2.8 MPa), a 25–30-fold reduction in mix design development time, and an 8–12% decrease in the production cost of B25-grade concrete while maintaining compliance with regulatory standards.
The system is intended to be implemented as a web-based application, “BetoNai KZ,” featuring a bilingual interface in Kazakh and Russian.
Abstract for an Interdisciplinary Audience
We have initiated the development of a computational system designed for the autonomous determination of optimal concrete mix compositions for construction enterprises in Kazakhstan. The system analyzes a large number of mix designs obtained from real industrial production, explicitly accounting for the variability in local material quality, climatic conditions, and key economic parameters.
By leveraging artificial intelligence techniques, the system identifies optimal combinations of cement, sand, coarse aggregate, and chemical admixtures that achieve the required mechanical performance while minimizing production costs and reducing environmental impact.
Preliminary evaluations conducted at five industrial facilities indicate that the system can generate optimized mix designs within a few hours rather than several days, while improving strength prediction accuracy and reducing costs by up to 12%.
The solution is intended to be deployed as a web-based application accessible via a standard browser and does not require specialized programming expertise from end users.
Keywords: artificial intelligence in construction materials science; deep learning; concrete mix design; multi-objective optimization; genetic algorithms; ready-mix concrete; Kazakhstan
Why This Is Important Now
The relevance of this research is driven by the ongoing industrialization of the construction sector, the implementation of large-scale government housing initiatives, and the steady growth of ready-mix concrete production in Kazakhstan. These trends necessitate the integration of digital technologies to enhance product quality while reducing production costs.
The proposed developments are aligned with the provisions of the Law of the Republic of Kazakhstan "On Informatization", which promotes digital transformation across economic sectors, as well as the anticipated Law of the Republic of Kazakhstan "On Artificial Intelligence", intended to establish a legal framework for the deployment and governance of artificial intelligence systems.
Thus, the evolving regulatory environment creates a supportive framework for the implementation of intelligent solutions in the production of construction materials.
Value Table
Item
What is novel
Theoretical Contribution
First, within the context of Kazakhstan, a hybrid machine learning paradigm—combining neural networks and genetic algorithms—has been proposed and integrated with economic optimization strategies to enable multi-objective design of ready-mix concrete compositions, while explicitly accounting for region-specific characteristics of raw material resources and climatic conditions.
Methodology
Currently, efforts are underway to develop and validate a comprehensive methodological framework aimed at the accelerated design of concrete mixtures. The proposed approach incorporates automated data acquisition from industrial production processes, advanced data preprocessing techniques, and multi-objective optimization—considering parameters such as mechanical performance, cost efficiency, and environmental sustainability—through the application of state-of-the-art artificial intelligence algorithms.
Practical Implementation
The proposed intelligent system has been applied under real industrial conditions at ready-mix concrete plants in Kazakhstan, demonstrating its practical feasibility and effectiveness. The implementation was carried out using production data and operational constraints typical of industrial environments, including variability in raw material properties and fluctuating climatic conditions.
The system enables automated generation of optimized concrete mix compositions within a few hours, significantly reducing the time required compared to conventional trial-and-error methods. By integrating predictive modeling with optimization algorithms, the system supports decision-making in selecting material proportions that meet required performance criteria while minimizing production costs and environmental impact.
Field application results indicate a substantial improvement in the accuracy of compressive strength prediction, along with a measurable reduction in material costs (up to 12%) without compromising compliance with regulatory standards. Additionally, the system facilitates rapid adaptation to changes in material quality, which is critical for maintaining consistent product performance in industrial practice.
The developed solution is implemented as a web-based application (“BetoNai KZ”), providing an accessible interface for engineers and plant operators. The platform does not require advanced programming skills and can be integrated into existing production workflows, thereby supporting the digital transformation of concrete production processes.
Scientific Community
The proposed study is of significant interest to the scientific community at the intersection of construction materials science, artificial intelligence, and industrial optimization. It contributes to the growing body of research on data-driven design of cementitious materials by demonstrating how advanced machine learning techniques can be effectively integrated with classical engineering principles.
For researchers in concrete technology, the study provides a scalable framework for modeling complex relationships between mix composition and performance characteristics under real industrial conditions. For the artificial intelligence community, it offers a domain-specific application of hybrid modeling approaches, combining deep neural networks with evolutionary optimization methods in a multi-objective setting.
Furthermore, the use of a large-scale industrial dataset enhances the reliability and applicability of the findings, addressing a common limitation of many existing studies based on laboratory-scale data. The proposed methodology also opens new avenues for interdisciplinary research, particularly in the areas of sustainable construction, digital twins of material systems, and intelligent decision-support systems.
The study is expected to stimulate further research on the integration of artificial intelligence into construction materials engineering, including the development of more interpretable models, incorporation of uncertainty quantification, and expansion toward lifecycle-based optimization of materials.
Three Key Questions
1. Who Will Benefit from This Work?
The proposed system is expected to deliver tangible benefits to multiple stakeholder groups. Ready-mix concrete plants and construction companies in Kazakhstan will be able to optimize production processes and reduce operational costs through data-driven mix design. Concrete technologists will benefit from a significant reduction in the time required to develop and validate new mix compositions.
Research institutions and universities will gain a robust platform for advancing studies in the digitalization of construction materials science, particularly in the application of artificial intelligence to material design. In addition, governmental bodies may utilize the system as a tool for monitoring the quality, consistency, and environmental performance of construction materials within the regulatory framework.
2. Where Can It Be Applied Now?
The proposed system can be readily deployed across multiple practical domains. It is directly applicable in ready-mix concrete production facilities in Astana, Almaty, Shymkent, and other urban centers in Kazakhstan, where it can support real-time optimization of mix design under variable production conditions.
In design and engineering firms, the system can be utilized for the development of technical specifications for concrete, enabling more accurate and performance-oriented material selection. Construction companies may apply the system as a decision-support tool for monitoring the quality and consistency of supplied concrete, thereby reducing risks associated with material variability.
Furthermore, the system can be integrated into academic curricula in civil engineering and construction materials programs, serving as a practical platform for training students in the application of digital technologies and artificial intelligence in materials science.
3. What Research Could Logically Extend This Study?
Several research directions can naturally build upon the proposed framework. First, the integration of the system with Internet of Things (IoT) sensors deployed on production lines would enable the development of a real-time self-learning system capable of continuous data acquisition and model updating.
Second, expanding the predictive capabilities of the system to include long-term performance indicators—such as durability and frost resistance—is essential for enhancing its applicability in harsh climatic conditions.
Third, the development of specialized modules tailored to advanced categories of concrete, including high-strength concrete, self-compacting concrete (SCC), and fiber-reinforced concrete, represents a critical step toward broadening the system’s industrial relevance.
Finally, the establishment of a national platform for anonymized data sharing among concrete producers would significantly improve model accuracy and robustness by enabling access to large-scale, diverse datasets, thereby fostering collaborative innovation across the industry.
Open Questions to the Scientific Community
How can a mechanism be established to facilitate voluntary sharing of anonymized industrial production data among competing enterprises in order to improve the accuracy of artificial intelligence models for concrete mix design, while simultaneously ensuring data confidentiality and preserving competitive advantages?
What are the ethical and legal implications of delegating final decision-making in concrete mix design to artificial intelligence systems, and how should responsibility be delineated between automated recommendations and human expert judgment?
ARTICLE
1. Introduction
The construction sector in Kazakhstan is currently experiencing substantial growth: the production of ready-mix concrete increased from 12.4 million m³ in 2018 to 18.7 million m³ in 2023. However, conventional concrete mix design methods—based on empirical correlations and iterative selection of components—remain time-consuming (typically requiring 5–7 days to develop a single mix) and do not consistently ensure an optimal balance between strength, cost efficiency, and environmental sustainability.
Specific characteristics of the Kazakhstani construction materials market introduce additional challenges. These include significant variability in the quality of locally sourced aggregates (with the coefficient of variation in compressive strength of crushed stone from different deposits reaching 18–22%), reliance on imported chemical admixtures, and a sharply continental climate with temperature fluctuations of up to 80°C.
These conditions necessitate specialized strategies to ensure frost resistance and long-term durability of concrete. Traditional design methodologies developed for stable conditions in developed countries do not adequately account for these region-specific factors.
Artificial intelligence and machine learning offer promising solutions to these challenges. A comprehensive review of the global scientific literature demonstrates the effective application of neural networks for predicting compressive strength of concrete [1–3], genetic algorithms for optimizing material compositions [4,5], and hybrid approaches for solving multi-objective optimization problems [6,7]. Nevertheless, a significant proportion of these studies rely on datasets derived from developed countries with standardized raw material quality, limiting their applicability to the conditions of developing economies.
The aim of this study is to establish a theoretical foundation and develop a functional intelligent system for the accelerated multi-objective design of ready-mix concrete compositions, specifically adapted to the unique characteristics of Kazakhstan’s raw material base, climatic conditions, and economic environment. The system is intended to simultaneously optimize three key parameters: achieving target compressive strength, minimizing production costs, and reducing carbon emissions.
2. Research Methodology
2.1. Database Development
A unique database of industrial ready-mix concrete compositions from Kazakhstani enterprises was established to train the intelligent system. Between January 2023 and December 2024, with the support of the Association of Ready-Mix Concrete Producers of Kazakhstan, data were collected from 12 plants located in Astana (4 plants), Almaty (3), Shymkent (2), Karaganda (2), and Aktobe (1).
The database comprises 2,847 industrial mix designs corresponding to concrete strength classes ranging from B15 to B40, each accompanied by comprehensive documentation. This includes precise proportions of all constituents (cement, fine aggregate, coarse aggregates of 5–10 mm, 10–20 mm, and 20–40 mm fractions, water, plasticizers, and air-entraining admixtures); compressive strength test results at 7, 14, and 28 days (three specimens per testing age); material properties according to quality certificates; curing conditions; and the cost of individual components in Kazakhstani tenge at the time of production.
Particular attention was given to collecting detailed information on the properties of locally sourced materials. For cement, the recorded parameters included grade, manufacturer, activity, standard consistency, and setting time. For aggregates, the dataset includes information on source deposit, strength, frost resistance, content of fine and clay particles, and particle size distribution. This approach enabled the incorporation of the high variability in raw material quality characteristic of the Kazakhstani construction materials market.
2.2. Intelligent System Architecture
The proposed intelligent system is structured as a three-tier architecture integrating modules for data preprocessing, property prediction, and multi-objective optimization (see Figure 1).
Level 1: Data Preprocessing and Feature Engineering At the initial stage, the input data are subjected to normalization procedures, while outliers are handled using the interquartile range (IQR) method. Derived features are subsequently generated, including the water-to-cement ratio, aggregate grading modulus, and volumetric cement concentration. In addition, categorical variables—such as cement type and aggregate source—are appropriately encoded to ensure compatibility with machine learning algorithms.
Level 2: Ensemble of Predictive Models
To predict 28-day compressive strength, a hybrid ensemble comprising three distinct model types was developed:
(a) a deep neural network (DNN) with an architecture of 15–128–64–32–1, employing ReLU activation functions and dropout regularization with a rate of 0.2;
(b) a gradient boosting model (XGBoost) consisting of 500 trees, with a maximum depth of 6 and a learning rate of 0.05;
(c) a support vector regression (SVR) model utilizing a radial basis function (RBF) kernel.
The final prediction is obtained as a weighted average of the individual model outputs, with weights of 0.5, 0.3, and 0.2, respectively. These weights are determined based on validation results, taking into account representative production conditions, including certified raw material properties, curing climate conditions, and component costs expressed in Kazakhstani tenge at the time of production.
Particular attention was devoted to the systematic collection of data on the properties of locally sourced materials. For cement, key parameters were carefully recorded, including grade, manufacturing plant, activity, standard consistency, and setting time. For aggregates, detailed information was documented on the source deposit, strength, frost resistance, content of fine and clay particles, and particle size distribution.
Level 3: Multi-Objective Optimization
To determine the optimal concrete mix composition, an enhanced version of the Non-dominated Sorting Genetic Algorithm II (NSGA-II) was employed. The algorithm was configured with a population size of 200 individuals and executed over 150 generations, incorporating adaptive crossover and mutation operators to improve convergence and solution diversity.
The optimization procedure simultaneously targets three objective functions:
f₁(x) = |fₜₐᵣ - fₚᵣₑ(x)| → min
f₂(x) = Σ(cᵢ · xᵢ) → min
f₃(x) = Eₖₒ₂(x) → min
where fₜₐᵣ — target compressive strength, fₚᵣₑ(x) — redicted Compressive Strength of Mix x by the Ensemble Model, cᵢ — cost of the iii-th component tenge per kilogram, xᵢ — Dosage of the i-th Component kg/m³³, Eₖₒ₂(x) — carbon footprint of the mix composition, expressed in kilograms of CO₂-equivalent per cubic meter (kg CO₂-eq/m³)
The system explicitly incorporates technological constraints in accordance with ST RK 1310-2005 and GOST 7473-2010: a water-to-cement ratio within the range of 0.35 ≤ W/C ≤ 0.70; cement content between 200 ≤ C ≤ 550 kg/m³; sand proportion in the aggregate mixture within 30% ≤ r ≤ 45%; and workability of the concrete mix classified as P2–P4 (slump ranging from 5 to 15 cm).
2.3. Model Training and Validation
The dataset was partitioned into training (70%, n = 1,993), validation (15%, n = 427), and test (15%, n = 427) subsets using stratified sampling based on concrete strength classes. To mitigate the risk of overfitting, regularization techniques were applied (L2 regularization for deep neural networks with α = 0.001), along with early stopping criteria; however, no improvement in validation performance was observed over 20 consecutive epochs.
The performance of the predictive models was evaluated using the following metrics: the coefficient of determination (R²), root mean square error (RMSE), and mean absolute percentage error (MAPE).
The training process was conducted on a computational cluster using TensorFlow 2.13, XGBoost 1.7, and scikit-learn 1.3 libraries within the Python 3.10 programming environment.
2.4. Industrial Validation
To assess the practical applicability of the proposed system, a comprehensive industrial validation was conducted at five independent ready-mix concrete plants that were not involved in the development of the training dataset. For each participating facility, the system generated optimized concrete mix designs corresponding to strength classes B20, B25, and B30.
These mixes were produced under real industrial conditions, and compressive strength tests were performed in accordance with GOST 10180-2012. The experimental results were subsequently compared with both the plants’ existing production benchmarks and the strength predictions generated by the system.
3. Results
3.1. Strength Prediction Accuracy
The ensemble of predictive models demonstrated high accuracy when applied to the test dataset, achieving R² = 0.942, RMSE = 2.78 MPa, and MAPE = 6.4%. In contrast, conventional regression approaches, particularly multiple linear regression, yielded significantly lower performance (R² = 0.741, RMSE = 5.83 MPa), corresponding to an approximately 27% reduction in prediction accuracy.
The most favorable results were observed for medium-strength concrete classes (B20–B30), with a MAPE of 5.2%. In comparison, slightly higher prediction errors were recorded for higher-strength concretes (B35–B40), with MAPE reaching 8.7%, which can be attributed to the relatively limited representation of such compositions in the training dataset.
A feature importance analysis using the SHAP (Shapley Additive Explanations) methodology revealed that the primary factors influencing compressive strength are as follows: cement content (contribution of 34%), water-to-cement ratio (28%), cement activity (15%), strength of coarse aggregate (12%), and the type and dosage of plasticizer (11%).
These findings are consistent with established physicochemical principles governing the structural formation of cementitious materials.
3.2. Multi-Objective Optimization Performance
The NSGA-II genetic algorithm delineates a Pareto front comprising 15–25 non-dominated solutions, representing different trade-offs between compressive strength, cost, and environmental sustainability. The algorithm operates within 4–6 hours on a standard server (Intel Xeon E5-2680 v4, 64 GB RAM), delivering performance that is 25–30 times faster than conventional iterative mix design approaches.
Table 1 presents a comparative analysis of mix compositions optimized by the proposed system against existing production mixes used for B25-grade concrete at one of the enterprises participating in the validation study.
Table 1 — Comparison of B25-Grade Concrete Mix Compositions
The optimized mix design enables a 10% reduction in cement consumption (equivalent to 38 kg/m³) while maintaining the required strength performance, resulting in cost savings of 1,530 KZT/m³. Considering an annual production volume of 85,000 m³, this corresponds to an estimated economic benefit of approximately 130 million KZT per year.
3.3. Industrial Validation Results
A total of 45 concrete mixes were produced and evaluated across five independent validation sites (three strength classes per site, with three mixes in each class). The observed compressive strength deviated from the model predictions by an average of 5.8%, with a standard deviation of 2.4%. This level of variability is consistent with the inherent variability of concrete and indicates a substantial improvement in prediction accuracy compared to conventional design methodologies, for which the average deviation was reported at 12.3%.
All 45 mix compositions complied with established regulatory standards in terms of compressive strength, workability, and frost resistance. The average cost reduction ranged from 8% to 12%, depending on the specific concrete class and the prevailing material price structure at each production facility.
The most significant cost savings (14.2%) were achieved at a plant in Karaganda for B20-grade concrete, primarily due to the optimization of aggregate gradation and a reduction in the use of high-cost plasticizers.
3.4. Carbon Footprint Reduction
Material optimization with explicit consideration of environmental criteria resulted in an average reduction of 7.2% in the specific carbon footprint of concrete, primarily due to decreased cement consumption, which is the dominant source of CO₂ emissions in concrete production.
For B25-grade concrete, emissions were reduced from 284 kg CO₂-eq/m³ to 263 kg CO₂-eq/m³. Considering that the annual production volume of ready-mix concrete in Kazakhstan is approximately 18.7 million m³, the potential emission reduction is estimated at up to 393,000 tonnes of CO₂-equivalent per year.
4. Discussion
The obtained results demonstrate that the integration of deep learning methodologies with evolutionary optimization algorithms establishes a qualitatively superior paradigm for concrete mix design compared to conventional approaches.
A key factor contributing to this improvement is the development of a representative industrial dataset that accurately captures the variability in raw material quality and the technological conditions characteristic of concrete production in Kazakhstan.
A comparative analysis with international studies indicates that the achieved prediction accuracy (R² = 0.942) is consistent with the performance reported in leading studies [1, 2, 6, 15]. Notably, however, the proposed system is among the first to be explicitly adapted to the conditions of a developing market characterized by pronounced variability in raw material quality.
Previous studies [3, 13, 14] have also employed multi-objective optimization; however, they have predominantly focused on binary trade-offs (e.g., strength–cost or strength–environmental impact). In contrast, the proposed system simultaneously optimizes three independent criteria, thereby providing a more comprehensive and practically relevant optimization framework.
The economic benefits of implementing the proposed system can be categorized into three primary components: immediate material cost savings (ranging from 8% to 12% of total production costs), a substantial reduction in the time required for developing new mix designs (from 5–7 days to 4–6 hours), and a decrease in defect rates due to more accurate prediction of material properties.
For a typical medium-sized production facility with an annual capacity of approximately 80,000 m³, the total projected economic benefit is estimated to range between 110 and 150 million KZT per year.
A significant advantage of the proposed system lies in the interpretability of its results, enabled by the application of explainable artificial intelligence (XAI) techniques, specifically SHAP (Shapley Additive Explanations).
This approach allows engineers not only to identify optimal mix compositions but also to understand the key factors influencing the model’s predictions. Consequently, this enhances trust in the system and facilitates its integration into industrial practice.
5. Where and Why the Proposed Approach May Not Be Applicable
The developed intelligent system demonstrates high effectiveness for conventional classes of ready-mix concrete; however, several limitations should be carefully considered when applying it in practice.
1. Specialized concrete types.
The system was trained on datasets derived from conventional normal-weight concrete classes ranging from B15 to B40. Its application to specialized concrete types—such as high-strength concrete (B45–B60), lightweight, cellular, fiber-reinforced, and self-compacting concrete—requires additional training on relevant datasets. The physicochemical properties of these materials may differ substantially, potentially leading to reduced prediction accuracy.
2. Non-standard components.
The current database includes the predominant types of cement used in Kazakhstan (primarily PC 400 and PC 500), along with commonly used aggregates and admixtures. The introduction of non-standard components—such as specialized cements, synthetic aggregates, or novel admixtures—may result in decreased prediction reliability due to their absence in the training data.
3. Extreme climatic conditions.
The models were developed based on data from enterprises located in regions with moderate and sharply continental climates. The application of the system to concrete intended for extreme environmental conditions (such as Arctic regions or arid desert climates) requires additional validation and, potentially, recalibration of the models.
4. Long-term performance.
The system is primarily designed to predict 28-day compressive strength and other short-term performance indicators. However, the assessment of long-term properties—such as durability, corrosion resistance, and shrinkage—requires additional modeling efforts and extensive longitudinal data, which are currently limited.
5. Level of digitalization at industrial facilities.
The effectiveness of the system is inherently dependent on the quality of input data. At facilities characterized by low levels of automation and inadequate raw material quality control—where data collection is sporadic or inaccurate—the reliability of the system’s predictions may be compromised. A baseline level of digitalization in production data management is therefore required to ensure optimal system performance.
6. Mini Dialogue with the Reader
Question: Why does the system not simply use the most efficient mix compositions from the database instead of generating new ones?
Answer: Each production facility operates with its own set of materials, including cement supplied from specific plants, aggregates sourced from particular geological deposits, and proprietary admixture systems. A mix design that is optimal for one facility may not be directly applicable to another due to differences in the physicochemical properties of raw materials.
The proposed system addresses this challenge by solving an optimization problem tailored to the specific materials and objectives of a given enterprise (e.g., cost minimization or reduction of carbon emissions). In this sense, the distinction is analogous to that between an off-the-shelf suit and a custom-tailored garment: while the former may be functional, it rarely provides an optimal fit.
Question: Can the system be used to audit and improve existing mix designs?
Answer: The system can effectively operate in an audit mode. Existing mix formulations can be input into the system, which then evaluates their performance and associated costs, while also identifying potential optimization opportunities.
For instance, the system may detect excessive cement consumption and recommend reductions of approximately 5–8% without compromising structural integrity. This capability is particularly valuable when raw material suppliers change or when material prices increase, as it enables efficient adjustment of mix designs in response to updated production conditions.
7. Use of Artificial Intelligence in the Study
In accordance with the principles of scientific transparency, the following section provides a comprehensive description of the application of artificial intelligence tools at various stages of the research.
Data preprocessing and analysis.
At the initial stages of data processing and refinement, automated outlier detection algorithms were employed using the Isolation Forest method from the scikit-learn library. This approach enabled the identification and removal of 127 records (4.5% of the original dataset) that exhibited significant measurement or documentation inconsistencies.
To address missing values related to aggregate properties, a k-nearest neighbors imputation method (kNN imputation, k=5k = 5k=5) was implemented.
Development of Predictive Models. The architecture of the deep neural network was optimized using Bayesian hyperparameter optimization implemented via the Optuna library. A total of 240 network configurations were evaluated, resulting in the selection of an optimal architecture aimed at minimizing the root mean square error (RMSE) on the validation dataset.
Model training was performed using the Adam optimization algorithm with an initial learning rate of 0.001, combined with an exponential decay schedule.
To assess the influence of individual features on the model predictions, the SHAP (Shapley Additive Explanations) library (version 0.42) was employed. Visualization tools were generated to illustrate feature importance and the relationships between SHAP values and corresponding feature values, enabling validation of the physical plausibility of the models and identification of nonlinear interactions.
Manuscript Preparation.
During the initial structuring of the manuscript and the drafting of preliminary versions of individual sections, the large language model GPT-4 developed by OpenAI was utilized. The model was used solely as a tool for generating preliminary text drafts, which were subsequently subjected to thorough scientific editing, refinement, and verification by the authors.
All concepts, conclusions, data interpretations, and final formulations presented in this study are the result of the authors’ intellectual contributions.
Graphical representations and diagrams were generated using the matplotlib 3.7 and seaborn 0.12 libraries within the Python programming environment. The selection of color schemes was performed in accordance with accessibility guidelines, ensuring compatibility for individuals with color vision deficiencies.
8. Conclusion
This study establishes a scientific foundation and presents the development of an intelligent system model for the multi-purpose design of ready-mix concrete compositions, specifically tailored to the characteristics of the construction sector in Kazakhstan. The key findings of the study are as follows:
A unique database comprising 2,847 industrial mix compositions from 12 Kazakhstani plants was compiled, including comprehensive documentation of local raw material properties. This dataset provides a robust foundation for further research in the digitalization of construction materials science.
2. A comprehensive ensemble of predictive models (DNN + XGBoost + SVR) was developed and rigorously validated, achieving high accuracy in predicting 28-day compressive strength (R² = 0.942), which represents a 27% improvement compared to conventional methodologies.
3. A multi-objective optimization framework based on the NSGA-II genetic algorithm was implemented, enabling the simultaneous optimization of compressive strength, production cost, and carbon emissions, and generating Pareto-optimal solutions.
Industrial validation at five independent facilities confirmed the practical viability of the system: mix design development time was reduced by a factor of 25–30, production costs decreased by 8–12%, and carbon emissions were reduced by 7.2%.
The developed BetoNai KZ system represents a comprehensive solution for the Commonwealth of Independent States (CIS) in the field of intelligent concrete mix design, taking into account regional specificities and delivering tangible economic and environmental benefits.
The system is intended for industrial deployment across enterprises in Kazakhstan and can be adapted for application in other countries of the region.
Future research directions include extending the system’s capabilities to predict durability and frost resistance, integrating it with Internet of Things (IoT)–based production management systems to enable the development of real-time self-learning models, and developing specialized modules for different types of concrete.
References
1. Mirzaei A., Aghsami A. A Hybrid Deep Reinforcement Learning Architecture for Optimizing Concrete Mix Design Through Precision Strength Prediction // Mathematical and Computational Applications. — 2025. — Vol. 30, No. 4. — Article 83. DOI: 10.3390/mca30040083
2. Chen F., Xu W., Wen Q., Zhang G., Xu L. Advancing Concrete Mix Proportion through Hybrid Intelligence: A Multi-Objective Optimization Approach // Materials. — 2023. — Vol. 16, No. 19. — Article 6448. DOI: 10.3390/ma16196448
3. Oviedo Carrascal A.I., Londoño J.M., Vargas J.F., Zuluaga C., Gómez A. Modeling and Optimization of Concrete Mixtures Using Machine Learning Estimators and Genetic Algorithms // Modelling. — 2024. — Vol. 5, No. 3. — P. 626-645. DOI: 10.3390/modelling5030034
4. Li Y., Li H., Li Y., Jin C. Optimum design of high-strength concrete mix proportion for crack resistance using artificial neural networks and genetic algorithm // Frontiers in Materials. — 2020. — Vol. 7. — Article 590661. DOI: 10.3389/FMATS.2020.590661
5. Tahmouresi A., Robati A., Urgessa G., Haghighi H. A Combined Genetic Algorithm-Artificial Neural Network Optimization Method for Mix Design of Self Consolidating Concrete // International Journal of Structural and Civil Engineering Research. — 2021. — Vol. 10, No. 3. — P. 106-112. DOI: 10.18178/IJSCER.10.3.106-112
6. Xu X., Zhang Y. Machine Learning the Concrete Compressive Strength From Mixture Proportions // ASME Open Journal of Engineering. — 2022. — Vol. 1. — Article 011038. DOI: 10.1115/1.4055194
7. Dias P.P., Jayasinghe L.B., Waldmann D. Machine learning in mix design of Miscanthus lightweight concrete // Construction and Building Materials. — 2021. — Vol. 302. — Article 124191. DOI: 10.1016/J.CONBUILDMAT.2021.124191
8. Ziolkowski P., Niedostatkiewicz M., Kang S.-B. Model-Based Adaptive Machine Learning Approach in Concrete Mix Design // Materials. — 2021. — Vol. 14, No. 7. — Article 1661. DOI: 10.3390/MA14071661
9. Mandal S., Shiuly A., Sau D., Mondal A.K., Sarkar K. Study on the use of different machine learning techniques for prediction of concrete properties from their mixture proportions // AI in Civil Engineering. — 2024. — Vol. 3, No. 8. DOI: 10.1007/s43503-024-00024-8
10. Zhang J., Li T., Yao Y., Hu X., Zuo Y. Optimization of mix proportion and strength prediction of magnesium phosphate cement-based composites based on machine learning // Construction and Building Materials. — 2024. — Vol. 411. — Article 134738. DOI: 10.1016/j.conbuildmat.2023.134738
11. Ren Q., Ding L., Dai X., Li C., Li C. Optimizing mix design of concrete with manufactured sand for low embodied carbon and desired strength using machine learning // Construction and Building Materials. — 2024. — Vol. 458. — Article 139407. DOI: 10.1016/j.conbuildmat.2024.139407
12. Tavares C., Wang X., Saha S.K., Grasley Z. Machine Learning-Based Mix Design Tools to Minimize Carbon Footprint and Cost of UHPC. Part 1: Efficient Data Collection and Modeling // Cleaner Materials. — 2022. — Vol. 6. — Article 100082. DOI: 10.1016/j.clema.2022.100082
13. Chen B., Zhou H., Xia D., Huang S.H. Big-data based concrete mix proportion optimization and application development // Proceedings of the 2nd International Conference on Concrete and Concrete Structures. — 2023. — P. 1-8. DOI: 10.4203/ccc.2.6.4
14. Shui Z., Xu Z., Gao X., Zhang S.Q. Multi-objective optimization of concrete mix design based on machine learning // Journal of Building Engineering. — 2023. — Vol. 78. — Article 107396. DOI: 10.1016/j.jobe.2023.107396
15. Song Y.N., Wang X., Li H., He Y., Zhang Z. Mixture Optimization of Cementitious Materials Using Machine Learning and Metaheuristic Algorithms: State of the Art and Future Prospects // Materials. — 2022. — Vol. 15, No. 21. — Article 7830. DOI: 10.3390/ma15217830