Deep learning techniques in concrete powder mix designing

Karam Ali Hadi; Aseel Sultan Ridha

doi:10.1515/eng-2022-0588

Article Open Access

Deep learning techniques in concrete powder mix designing

Karam Ali Hadi and Aseel Sultan Ridha

Published/Copyright: March 8, 2024

Published by

Become an author with De Gruyter Brill

Submit Manuscript Author Information

From the journal Open Engineering Volume 14 Issue 1

Abstract

The water-cement ratio is multiple phases procedure in which we aim to determine the most optimal combination for producing high-performing concrete. In modern literature and business practice, there are various methods for designing concrete mixes, although the Three Equation Method-inspired procedures are by far the most widely used. Concrete compressive strength is one of the fundamental properties that determines its class. Foreseeable compressive strength concrete is necessary to promote the use of concrete structures. The primary feature of its durability and safety is. Deep learning has recently received a lot of attention, and the prospects for this technology are even brighter. Machine learning algorithms have advanced to the point that they can recognize patterns, which are difficult for humans to recognize. This has sparked interest in data mining on enormous datasets. In this research, we aim to utilize cutting-edge developments in machine learning techniques for the production of concrete mixes. To provide the ideal structure of a synthetic neural network that has been chosen, we compiled a comprehensive dataset of concrete mixtures, complete with laboratory destructive test results. A mathematical formula that may be used in practical applications has been developed from the creation of an artificial neural network.

Keywords: deep learning techniques; concrete powder mix; water-cement ratio; concrete compressive strength

1 Introduction

“Concrete mix design” is a significant but mysterious topic that demands an in-depth comprehension of several specialized concerns. The structure can be used confidently if concrete with the right strength and other utility criteria is obtained. The strengthening and hydration of concrete are permanent procedures. As a result, any flaws in the design of the mixture of concrete are extremely expensive to the owner during construction and reduce the profitability of the edifice because of its diminished durability. To improve properties such as concrete strength, density, workability, or durability, concrete mixtures of cement, coarse and fine aggregate, and water are generally reinforced with additives and admixtures

Concrete is produced as the final product from a concrete mixture. The cement hydration process, and cement and water chemical thermal interaction initiates the concrete strengthening procedure. Cement hydration producers are gel, hydroxide, and a few secondary chemicals that aid in fine and coarse aggregate bonding. Throughout the hydration procedure, the by-products of hydration progressively settle on initial cement particles and cover the area vacated by water. When the water molecules retreat or there is no longer any unreacted cement, the procedure of hydration is complete. Concrete hardens furthermore and reaches full compressive strength during the 28 days [1,2,3]. Concrete mix design is about choosing the right quantities of cement, fine and coarse aggregate, and water to generate concrete with the desired qualities [4,5,6]. The design of concrete mixtures is progressing at a steady pace. The most common approach for measuring the number of primary ingredients required has been utilized for years and includes evaluating the bending strength of concrete mortar [7,8,9,10]. These systems have numerous drawbacks and are time-consuming to implement. We want to demonstrate a way to create concrete using a mathematical formula derived by a machine learning algorithm. The adopted neural network architecture is described in the accompanying paper, which will be fed by an extensive collection of concrete mixture dataset. We offer a mathematical equation as a conclusion, for calculating concrete compressive strength. The created method will use cement, water, and fine and coarse aggregates as its four input parameters to determine the compressive strength of concrete.

The proposed formula involves boundary conditions and does not precisely capture the behavior of concrete. However, it is a first step toward using machine learning techniques in the concrete mix design. It can be used to make a preliminary estimate of the concrete class in its current state. In future attempts, we plan to focus on concrete mixtures design in concern with the aspects of durability and estimation of service lifetime. The use of concrete admixtures such as superplasticizers, for example, would be essential.

2 Concrete mix designing based on deep learning techniques

2.1 Mathematical model

Determining the correct quantitative content and percentage of concrete mixture elements is the main purpose of water cement development. We must select a combination that permits us to get the most precise results. Concrete performance is defined by various characteristics, the most important of which are compressive strength and durability. In the concrete mix design, both strength and durability should be considered. In an aggressive setting, the question of durability is critical [11,12,13,14,15]. According to our research, there are a few popular techniques to build a mixture of concrete within European corporate engineering practice. Three of these methods are the Bukowski, Eyman and Klaus, and Paszkowski methods. Following results are found through the “Three Equations Method,” also known as the Bolomey method, a mixed experimental-analytical technique [16,17]. It implies that the experimental evidence should back up the mathematical technique. We use analytical techniques to calculate the volume of required components and destructive laboratory testing to confirm the results.

To determine the three required values, we utilize a fundamental endurance, consistency, and stiffness equation: the amount of water, cement, and aggregate represented in kg/m³. The first formula (equation (1)) is the compressive strength formula, also known as the Bolomey formula.

(1) f cm = A 1 , 2 ( C / W ± 0.5 ) ( MPa ) ,

where f _cm is the concrete’s medium compressive strength in N/mm². A _1,2 denotes coefficients that vary based on cement grade and aggregate type, C is the amount of cement in one cubic meter of concrete in kg, W denotes the quantity of water in one cubic meter concrete in kg.

Equation (2), consistency equation, is incorporated into the watershed management formula to make a geopolymer concrete with the desired texture.

(2) W = C w c + K w k ,

where W is the weight of water in one cubic meter of concrete in kg, C is the weight of cement in one cubic meter of concrete in kg, K is the aggregate water demand index in dm³/kg, w _c is the cement-water demand index.

The simple volume formula includes equation (3).

(3) C / μ c + K / μ k + W = 1 , 000 [ d m 3 ] ,

where W denotes the water quantity in one cubic meter of concrete in kg, C denotes the quantity of cement in one cubic meter of concrete in kg. K denotes the quantity of aggregate in one cubic meter of concrete in kg, µ _c denotes density of cement in kg/dm³, and µ _k: denotes the aggregate density per dm³ in kg.

2.2 CNN modeling

Deep learning has been a rapidly increasing field of expertise in recent years. This technology is a branch of artificial intelligence science that includes subjects like statistics, computer science, and robotics [18,19,20,21]. In practice, machine learning tries to combine numerous revolutionary computer science breakthroughs to develop a system which can learn from datasets and, as a result, search themes and connections between variables and sets of variables that would be difficult to discover using conventional methods. In this scenario, learning can be thought of as implementing a complex algorithm. Convolution neural networks (CNNs) are one of the most common machine learning algorithms. Beginning with the initial input data, each composing module in CNN turns what is represented at a particular level into a higher and more complex level, similar to how a regular deep learning neural network works. Natural properties or complex functions could be learned by composing enough of these modifications [22,23,24]. CNN training is a comprehensive learning method [25,26,27,28] that may implicitly learn characteristics from data. As a result, manually extracting data features is unnecessary, as is initial processing or rebuilding the initial information [29]. The essential components of the first few units of CNN design are extremely similar. They use a serial convolution layer and a pooling layer to arrange data features layer-by-layer, and CNN was called after this architecture. The final unit comprises a few completely interconnected layers and a classic classification model. In many practical applications, recollecting data or rebuilding models is expensive, if not impossible, using most classic machine learning approaches [30]. Current CNN models (shown in Figure 1) demand a lot of processing power and have complicated computational requirements. Transfer learning is an excellent choice since they are exposed to local optimization difficulties or overfitting [30,31,32,33]. Another benefit of transfer learning is that it does not necessitate a huge amount of data records; however, it can achieve improved accuracy with a smaller dataset.

Figure 1

Convolution neural networks.

3 Simulation setup and results

Deep learning technique is one of the most used concrete mix designing algorithms. One of the most artificial expert system that is proposed is a CNN (shown in Figure 2) that can estimate the compressive strength of a concrete mix based on a huge number of tested concrete mix mixtures.

Figure 2

Exact CNNs structure.

The CNN calculates the concrete’s strength according to the proportions of the four essential ingredients in a concrete mix: cement, fine and coarse aggregate, and water. We converted the built CNN accordingly and reduced it to a single equation, defining concrete’s 28 days strength as a function of the four factors. The equation can calculate concrete compressive strength and validate the concrete mix recipe (Figure 3).

Figure 3

Illustrated diagram for cement mixing system [15].

Setting a border constraint for this procedure seems fair. However, because the CNN was trained on a few samples, predicting how it will react to material concentrations outside of the specified limits may be difficult. The water-cement ratio must be properly controlled since the right balance is required for full hydration of the cement. The impact of plasticizers has not been investigated. All algorithmic steps for classification and predication are shown in Figure 4.

Figure 4

The grading curves.

Many aspects, such as the curing procedure, indirectly affect the produced concrete strength and were not considered in the analysis. We anticipated that strict quality control would provide full-strength concrete. The most adopted database generation are listed in Table 1.

Table 1

Database adopted generation

Compressive strength after 28days	Cement	Water	Sand 0–2 mm
cs_28target	Cement input	Water input	Fine_aggregate input
The compressive strength of concrete at 28 days after hydration Is considered as full strength	The weight of cement added to the mixture	The weight of water added to the mixture	The weight of sand added to the mixture

Table 2 shows each input variable’s minimum, maximum, and average values.

Table 2

Input features range

Input features	Minimum (kg/m³)	Maximum (kg/m³)	Average (kg/m³)
Cement	86.00	540.00	278.00
Water	121.80	247.00	182.42
Fine aggregate (sand 0–2 mm	372.00	1329.00	768.55
Coarse aggregate (aggregate above 2 mm	597.00	1490.00	969.408

The parameters in Table 2 were separated into inputs and targets, that describe variables for input and output, respectively. Concrete strength gradually increases to full strength after starting the cement hydration process. During our deliberations, we assumed that concrete would achieve its intended compressive strength after 28 days. The concrete has some strength before the 28 days but cannot be deemed as full strength.

In our investigation, we thought the concrete attained full strength because the mixture was designed for it. The examined mixtures are presented in Table 1. The grading and aligning curves for the designed mixtures are shown in Figure 4.

The main algorithmic steps for implementation of the proposed system is shown in Figure 5, where the CNN is trained first with the training samples.

Figure 5

Flowchart for the proposed system.

Figure 6 shows the machine learning accuracy in the classification and prediction of the cement ratio system, whereas Figure 7 shows the comparison of the overfitting cases.

Figure 6

Accuracy of classification and prediction of cement ratio system.

Figure 7

Comparison of the overfitting cases.

Table 3 demonstrates how the CNN accuracy is affected by the input and hidden layer structures with three phases training, validation, and testing phase.

Table 3

Models’ metric performance

Average		Total duration	Accuracy%	loss
Training	Total	1,028	89.72	0.23
	Transfer learning	792	93.12	0.18
	Random initialization	1,264	86.32	0.31
Validation	Total	—	84.16	0.35
	Transfer learning	—	88.61	0.27
	Random initialization	—	79.72	0.47
Test	Total	6.87	87.42	0.37
	Transfer learning	6.59	89.52	0.21
	Random initialization	7.14	85.34	0.52

Figure 3. Schematic diagram of the cement mixing system.

The statistics shown above suggest that for datasets with a small number of pictures, simple networks with few parameters and shallow depth can achieve excellent accuracy and efficiency. On the contrary, the use of complicated networks is prone to overfitting due to a lack of data, which would impair the training impact. Table 4 depicts CNN's performance, and it can be observed that it is nearly optimal for training data. Almost all of the datapoints are inside the 10% error limit. The CNN metamodel is shown to be more accurate for higher CS values than for CS values less than 50 MPa. The CNN on training and testing is seen to be 99 and 97%, respectively.

Table 4

Comparison results with literature research

No. 1	reference	Methods (accuracy 100%)	MSE	Max. error
1	[14]	AdaBoost regression 78%	96.84	20.44
2	[15]	Multi-layer perceptron 90%	46.82	20.84
3	[17]	Decision tree regression 94	26.09	21.20
4	Our proposed system	CNN 98.6%	10.06	10.7

4 Conclusion

The CNN formula showed low resilience for high strength concrete mixes (50 MPa and more). This could be owing to the limited amount of mixes used to train the CNN for these ranges. CNN’s behavior could indicate underfitting. We must emphasize that the method provided here is simply an introduction to the machine learning large application in concrete mix creation and does not cover the entire subject. It ignores certain critical concerns, such as the technological process and durability. Our research focuses on using machine learning in concrete mix ration and developing a practical tool for use in engineering practice. We created the best CNN architecture for the study and gave it a huge database of concrete mix formulas. A destructive laboratory test is associated with each concrete mix recipe record. The purpose of producing concrete with specific compressive strength is achieved by predicting optimal mixture of concrete materials using a neural network. More specifically, what materials ratio should be chosen to generate concrete with a suitable compressive value. Our database has 941 records.

Funding information: The manuscript was done depending on the personal effort of the author, and there is no funding effort from any side or organization.
Conflict of interest: There is no conflict of interest with anyone related to the subject of the manuscript or any competing interest.
Data availability statement: Most datasets generated and analyzed in this study are in this submitted manuscript. The other datasets are available on reasonable request from the corresponding author with the attached information.

References

[1] Marchon D, Flatt RJ. Mechanisms of cement hydration. In Science and Technology of Concrete Admixtures. Amsterdam, The Netherlands: Elsevier; 2016. p. 129–45.10.1016/B978-0-08-100693-1.00008-4Search in Google Scholar

[2] Scrivener K, Snellings R, Lothenbach B. A Practical Guide to Microstructural Analysis of Cementitious Materials. Boca Raton, FL, USA: CRC Press; 2018.10.1201/b19074Search in Google Scholar

[3] Kurdowski W. Cement and Concrete Chemistry. Berlin, Germany: Springer Science & Business; 2014.10.1007/978-94-007-7945-7Search in Google Scholar

[4] Young JF, Mindess S, Darwin D. Concrete. Upper Saddle River, NJ, USA: Prentice Hall; 2002.Search in Google Scholar

[5] Hover KC. The influence of water on the performance of concrete construction build. Mater. 2011;25:3003–13.10.1016/j.conbuildmat.2011.01.010Search in Google Scholar

[6] Jensen OM, Hansen PF. Water-entrained cement-based materials, I. Principles and theoretical background. Cem Concr Res. 2001;31:647–54.10.1016/S0008-8846(01)00463-XSearch in Google Scholar

[7] Jensen OM, Hansen PF. Water-entrained cement-based materials, II. Experimental observations. Cem Concr Res. 2002;32:973–8.10.1016/S0008-8846(02)00737-8Search in Google Scholar

[8] Jiménez Fernández CG, Barra BM, Valls del Barrio S, Aponte Hernández DF, Vázquez RE. Durability of recycled aggregate concrete designed with the equivalent mortar volume (EMV) method: Validation under the Spanish context and its adaptation to Bolomey methodology. Mater Construcción. 2014;64(313):e006.10.3989/mc.2013.00913Search in Google Scholar

[9] Abdelgader HS. How to design concrete produced by a two-stage concreting method. Cem Concr Res. 1999;29:331–7.10.1016/S0008-8846(98)00215-4Search in Google Scholar

[10] Abdelgader HS, Suleiman RE, El-Baden AS, Fahema AH, Angelescu N. Concrete mix proportioning using three equations method (Laboratory Study). Proceedings of the UKIERI Concrete Congress Innovations in Concrete Construction. Jalandhar, India: March 2013. p. 20–7.Search in Google Scholar

[11] Cartuxo F, De Brito J, Evangelista L, Jiménez RJ, Ledesma FE. Increased durability of concrete made with fine recycled concrete aggregates using superplasticizers. Materials. 2016;9(2):98.10.3390/ma9020098Search in Google Scholar PubMed PubMed Central

[12] Serralheiro MI, de Brito J, Silva A. Methodology for service life prediction of architectural concrete facades. Constr Build Mater. 2017;133:261–74.10.1016/j.conbuildmat.2016.12.079Search in Google Scholar

[13] Plank J, Sakai E, Miao CW, Yu C, Hong JX. Chemical admixtures—Chemistry, applications and their impact on concrete microstructure and durability. Cem Concr Res. 2015;78:81–99.10.1016/j.cemconres.2015.05.016Search in Google Scholar

[14] Huang H, Qian C, Zhao F, Qu J, Guo J, Danzinger M. Improvement on the microstructure of concrete by polycarboxylate superplasticizer (PCE) and its influence on durability of concrete. Constr Build Mater. 2016;110:293–9.10.1016/j.conbuildmat.2016.02.041Search in Google Scholar

[15] Arredondo-Rea SP, Corral-Higuera R, Gómez-Soberón JM, Gámez-García DC, Bernal-Camacho JM, Rosas-Casarez CA et al. Durability parameters of reinforced recycled aggregate concrete: Case study. Appl Sci. 2019;9:617.10.3390/app9040617Search in Google Scholar

[16] Andrade C, Gulikers J, Marie-Victoire E. Service Life and Durability of Reinforced Concrete Structures. Vol. 17, Berlin, Germany: Springer; 201810.1007/978-3-319-90236-4Search in Google Scholar

[17] Aïtcin PC. The durability characteristics of high performance concrete: A review. Cem Concr Compos. 2003;25:409–20.10.1016/S0958-9465(02)00081-1Search in Google Scholar

[18] Ahmad S. Reinforcement corrosion in concrete structures, its monitoring and service life prediction – A review. Cem Concr Compos. 2003;25:459–71.10.1016/S0958-9465(02)00086-0Search in Google Scholar

[19] Schabowicz K, Gorzela´nczyk T, Szymków M. Identification of the degree of degradation of fibre-cement boards exposed to fire by means of the acoustic emission method and artificial neural networks. Materials. 2019;12:656.10.3390/ma12040656Search in Google Scholar PubMed PubMed Central

[20] Godycki-Cwirko T, Nagrodzka-Godycka K, Wojdak R. Reinforced concrete thin-wall dome after eighty years of operation in a marine climate environment. Struct Concr. 2016;17:710–7.10.1002/suco.201500180Search in Google Scholar

[21] Schabowicz K, Gorzelánczyk T, Szymków M. Identification of the degree of fibre-cement boards degradation under the influence of high temperature. Autom Constr. 2019;101:190–8.10.1016/j.autcon.2019.01.021Search in Google Scholar

[22] Argiz C, Moragues A, Menéndez E. Use of ground coal bottom ash as cement constituent in concretes exposed to chloride environments. J Clean Prod. 2018;170:25–33.10.1016/j.jclepro.2017.09.117Search in Google Scholar

[23] Rajamane NP, Ambily PS, Nataraja MC, Das L. Discussion: Modified Bolomey equation for strength of lightweight concretes containing fly ash aggregates. Mag Concr Res. 2014;66:1286–88.10.1680/macr.14.00154Search in Google Scholar

[24] Zhang X, Deng S, Deng X, Qin Y. Experimental research on regression coefficients in recycled concrete Bolomey formula. J Cent South Univ Technol. 2007;14:314–7.10.1007/s11771-007-0272-0Search in Google Scholar

[25] Abdelgader HS, Saud AF, Othman AM, Fahema AH, El-Baden AS. Concrete mix design using the double-coating method. Betonw Fert Plant Precast Technol. 2014;80:66–74.Search in Google Scholar

[26] Rumman R, Bose B, Emon MAB, Manzur T, Rahman MM. An experimental study: Strength prediction model and statistical analysis of concrete mix design. Proceedings of the International Conference on Advances in Civil Engineering. Istanbul, Turkey: September 2016. p. 21–3.Search in Google Scholar

[27] Pereira FC, Borysov SS. Machine learning fundamentals. In: Antoniou C, Dimitriou L, Pereira F, editors. In Pereira Big Data and Transport Analytics. Ch. 2, Amsterdam, The Netherlands: Elsevier; 2019. p. 9–29.10.1016/B978-0-12-812970-8.00002-6Search in Google Scholar

[28] Shobha G, Rangaswamy S. Machine Learning. In: Gudivada VN, Rao CR, editors. In Computational Analysis and Understanding of Natural Languages: Principles, Methods and Applications. Vol. 38, Amsterdam, The Netherlands: Elsevier, 2018. Ch. 8; p. 197–228. 10.1016/bs.host.2018.07.004Search in Google Scholar

[29] Shrivastava VK, Pradhan MK, Minz S, Thakur MP. Rice plant disease classification using transfer learning of deep convolution neural network. Int Arch Photogram Remote Sens Spatial Inf Sci. Jul. 2019;XLII-3/W6:631–5.10.5194/isprs-archives-XLII-3-W6-631-2019Search in Google Scholar

[30] Hinton GE. Reducing the dimensionality of data with neural networks. Science. Jul. 2006;313(5786):504–7.10.1126/science.1127647Search in Google Scholar PubMed

[31] Gu J, Wang Z, Kuen J, Ma L, Shahroudy A, Shuai B, et al. Recent advances in convolutional neural networks. Pattern Recognit. May 2018;77:354–77.10.1016/j.patcog.2017.10.013Search in Google Scholar

[32] Bidkar K, Jadhao P. Prediction of strength of remixed concrete by application of orthogonal decomposition, neural analysis and regression analysis. Open Eng. 2019;9(1):434–43.10.1515/eng-2019-0053Search in Google Scholar

[33] Piotrowski T, Tefelski D, Polański A, Skubalski J Monte Carlo simulations for optimization of neutron shielding concrete. Open Eng 2012;2(2):296–03.10.2478/s13531-011-0063-0Search in Google Scholar

Received: 2023-06-08

Revised: 2023-12-27

Accepted: 2024-01-09

Published Online: 2024-03-08

This work is licensed under the Creative Commons Attribution 4.0 International License.

Articles in the same Issue

https://doi.org/10.1515/eng-2022-0588

Keywords for this article

deep learning techniques; concrete powder mix; water-cement ratio; concrete compressive strength

Creative Commons

BY 4.0