Combining Kriging meta models with U-function and K-Means clustering for prediction of fracture energy of concrete

https://doi.org/10.1016/j.jobe.2020.102050Get rights and content

Highlights

  • The accuracy and efficiency of the Kriging method has been investigated to predict the fracture energy of concrete.

  • The use of U-function and K-Means Clustering has improved the prediction results and reduced the average error.

  • The use of the proposed models requires less training data compared to the ANN.

  • The proposed models are more accurate and powerful than the ANN and the previous relationships.

Abstract

In this study, a combination of Kriging surrogate method with U-learning function and K-means clustering were used to predict the concrete fracture energy (as an output parameter) based on previous experimental data sets including compressive strength, maximum aggregate size and the water to cement ratio (as the input parameters). Therefore a collection of 246 data series obtained from previous studies was collected. The strength, accuracy, and efficiency of the proposed models were examined by selecting 10%, 30%, and 50% of the data for learning, and the results were compared with the previous equations. The results show that combining the Kriging method with the U-learning function in the work of fracture method (WFM) will increase the predictive power of fracture energy compared to basic Kriging and K-means clustering methods, and the previous relationships. However, the size effect method (SEM), the models created using K-means and 50% of the data has led to better forecasting results than other models. The value of the correlation coefficient (R2) of the proposed Kriging combination models and previous existing relationships are in the range of 0.59–0.95 and 0.14–0.69, respectively. The results show that the combination of the Kriging method, the U-learning function, and K-means clustering will reduce the time and cost of the experiments, as well as increasing the accuracy of concrete fracture energy prediction results using a small number of previous experimental data.

Introduction

The growth and expansion of micro-cracks and their transformation into large and continuous cracks will lead to the fracture of concrete [1]. Investigating the solutions to prevent and control the growth of micro-cracks in concrete elements increases safety and reduces the economic losses and casualties caused by the fracture and collapse [2]. As micro-cracks cover a large area and prevent the rapid expansion of cracks, structural engineers consider the size effect parameter (fracture energy (Gf) (initial fracture energy)) to examine the impact of the member's dimensions on its ultimate strength [3]. Therefore, when the purpose is to examine the performance of the real elements using small-scale samples in a laboratory, the significance of the size effect concept becomes more apparent. In addition, the classification of materials (brittle, quasi-brittle, and elastic-plastic) and the determination of their linear or nonlinear performance in structures are based on the size and recognition of the fracture process zone (FPZ) at the end of the crack [4,5]. To study the fracture parameters of different materials, use can be made of the size effect method [4,6], work fracture method [4,7], crack band method [8], effective crack method [9], two parameter model [10], cohesive crack method or fictitious crack method [11,12], double-G fracture method [13], KR-curve method [14,15], double-K fracture method [[16], [17], [18], [19]], boundary effect method and simplified boundary effect method [20,21].

In calculating the size effect fracture energy (Gf), the test specimens have the same geometry (shape), but vary in size. For this reason, the fracture energy calculated in size effect method (SEM) does not depend on the size, shape and geometry of the samples. As the maximum applied load and the area under the load-displacement graph before the maximum load is used to calculate the Gf, it is called the initial fracture energy (Fig. 1) [5]. Another method used to study the fracture energy is called work fracture method (WFM). In this method, the specific fracture energy (GF) is the same energy used to create the unit area of crack. Conventional-sized samples are used in WFM to calculate the GF, and the amount of the fracture energy depends entirely on the size and shape of the samples. In this method the fracture energy will be calculated by calculating the total area under the load-displacement diagram (Fig. 1), [5].

The fracture energy in WFM is always greater than the fracture energy obtained from SEM [22,23]. There are different points of view to the existence or non-existence of a clear relationship between these two parameters [24,25]. This difference is caused by various reasons such as the size and shape of the samples, how the experiments were performed, the distribution and the uncertainty of the fracture energy values obtained from the WFM after the maximum load and in the final part of the load-displacement diagram compared to the SEM fracture energy [4,23]. Bazant and Becq-giradun [24] reported a value of about 2.5 for the GF/Gf in typical concrete. While Beygi et al. [10] claimed that for self-compacting concrete (SCC), this ratio is 3.11. Ghasemi et al. [27] reported GF/Gf ratio for SCC with steel fibers to be 9.66. Researches by Kazemi et al. [28] showed that in high-strength concrete with steel fibers, the GF/Gf value is approximately equal to 10.5. The WFM method yields better fracture energy results than SEM because it considers the total area under the load-displacement diagram and the post cracking behavior of concrete [4]. Furthermore, the physical sample size is not adequate to describe the size effect on the fracture properties observed on laboratory-size samples because depending on the relationship between the sample width and the ligament transition length, a boundary effect and the size effect can exist [21].

Water-to-cement ratio (W/C), aggregate properties, amount and characteristics of admixtures, etc. are among the parameters affecting crack propagation and accelerating the fracture phenomenon in concrete [29]. According to research by Wittmann et al. [30], increasing the W/C in conventional concrete reduces the fracture energy. Carpinteri and Brighenti [31] reported that using the optimal amount of W/C generates the maximum fracture energy due to increased quality and strength of cement paste, especially in the area of contact with aggregates. They believed that this happens because the high strength of the cement paste almost prevents the spread of micro-cracks in this area. According to the studies by Nallathambi et al. [32], when the amount of W/C is high, the expansion and growth of cracks occurs more rapidly. Previous studies reported a decrease in fracture energy when W/C is too low or too high [26]. However, some researchers have presented different results on the relationship between W/C ratio and fracture parameters. For example, Mindess et al. [33] and Phillips and Binsheng [34] reported that the fracture energy does not change with increasing or decreasing the W/C ratio. According to Zhao et al. [35], there is no clear and tangible relationship between these two parameters. However, Ince and Alyamac [36] used Abrams’ law to confirm the existence of a specific relationship between W/C ratio and fracture characteristics.

Aggregates make up more than 70% of the concrete volume. For this reason, the performance and behavior of concrete at different ages are strongly influenced by the properties of these materials [37]. Although there is no specific relationship between the size of the aggregate and fracture toughness [38], increasing the amount and maximum size of aggregates (dmax) will have a positive effect on the failure properties [39,40]. Many researchers have reported that increasing the maximum size of the aggregates increases the fracture energy of concrete [[41], [42], [43]]. The studies by Petersson [44] showed that improving the quality and strength of the aggregates will have a positive effect on the fracture characteristics. Ince and Alyamac [36] reported that when the compressive strength of typical concrete (normal vibrating concrete) increases, the characteristics of the failure phenomenon such as energy and toughness increase. Compressive strength plays an important role in increasing fracture energy and toughness of high strength concretes. Although increasing the maximum size of the aggregates in this type of concrete will increase the compressive strength, unlike ordinary concretes, the failure energy of high-strength concretes is more affected by the type of aggregates [[32], [33], [34]].

Providing prediction models that have the desired strength, accuracy, and low cost to estimate the various parameters of materials, based on existing experimental input and output data sets, has always been interesting for researchers. In addition to multiple linear regressions, artificial intelligence methods such as artificial neural networks (ANN), adaptive neuro-fuzzy inference systems (ANFIS), and fuzzy and neuro-fuzzy systems have been used to estimate the complex performance of the structural materials [23,45,46]. Some Meta models such as Kriging surrogate, response surface methodology, dimension reduction prediction, etc. have also been used for this purpose [[47], [48], [49]]. Because concrete falls into the category of quasi-brittle materials and has a complex nonlinear behavior in fracture, assessment and obtaining an accurate fracture behavior model for this material play a key role in for an ideal, optimal, and safe design. Due to the cost and time reduction, it is very useful to provide an accurate model of failure behavior using existing experimental data sets and based on prediction methods and formulas [23].

Section snippets

Significance of the study

Investigating the concrete characteristics requires time, money, designs of experiments (DoE), preparation of materials and testing with special equipment at the desired ages. Therefore, the use of solutions that prevents or reduces the loss of cost, time and other disadvantages is needed in the concrete industry. In addition, human error due to the lack of skills and adequate knowledge of basic theories, the poor performance of the researcher during the experiment and equipment errors such as

Existing relationships to predict concrete fracture energy

Various equations and relationships have been proposed in recent decades to estimate the fracture energy of typical concrete using regression and the experimental samples. Some of these equations that have been used by researchers and codes are shown in Table 1. These relationships, which are often based on statistics and regression, are not accurate enough. Therefore surrogate methods are needed for more accurate assessments.

Kriging method

The original idea for the Kriging surrogate was developed by Daniel Krige and very fast became a common method as a low-cost simulation approach [60]. Kriging method is widely used in reliability assessment and collapse probability problems. Compared to the other surrogate methods, this method is known as an accurate interpolation method and the design points predicted by this method are exactly the same as the actual output values. The user can also increase the prediction accuracy of the

U learning function

In the past, the proposed Kriging models were presented inactive or fixed with the introduction of many points based on DoE concepts [69]. But Kriging's ability to determine the variance of the estimated data makes it possible to improve the DoE points with the help of an adaptive structure. To increase the accuracy of result prediction, when there is data with high variance, that point will be added to the DoE data set [48]. It is difficult to achieve the desired level of accuracy and

K-means clustering

The unsupervised classification process (clustering) is used to classify and place the data with similar characteristics in the same groups [73,74]. K-means clustering is widely used for data sets classification in K cluster. Despite the simplicity and power of discovering the hidden patterns of the data set, as well as the very good convergence rate of K-means clustering in achieving the optimal local point, the use of this method is limited. These limitations are due to introducing the number

Test data and input and output parameters

A data set collected by Nikbin et al. [23] where used in this study to predict the fracture energy of the concrete as an output parameter. A total of 246 data series where used. In order to summarize the information, the values of the input and output parameters are shown as domains in Table 2. Table 3 shows the details of each parameter. Many researchers have investigated the effects of various parameters such as aggregate type, aggregate content, aggregate size distribution, curing

Methodology

In prediction methods such as neural networks and ANFIS, determining a certain percentage of data for better network learning and providing an accurate model for estimating the engineering properties and materials is a big problem and in most cases, a lot of basic data sets are needed. In this study, the Kriging method, U-learning function and K-means clustering where combined to assess models that show the best performance in concrete fracture energy estimation with the lowest percentage of

Base kriging models

Fig. 3 shows the Gf and GF results of the base Kriging models versus experimental results. The results of the statistical parameters are given in Table 6 and Table 7 to examine the accuracy of the basic Kriging models. According to Table 6, Table 7, considering the aggregate shape, the WFM and SEM fracture energy prediction for all basic Kriging models have less error compared to the previous available relationships. However, predicting the fracture energy in the WFM using the basic Kriging

Conclusions

In this study, a database of 246 data sets was collected from past studies in which compressive strength, water-to-cement ratio and maximum aggregate size are considered as input parameters and the fracture energy as the output. In order to predict the fracture energy parameter of concrete, Kriging surrogate method was used by selecting 10%, 30%, and 50% of the data for training and combining this method with U-learning function and K-Means clustering. After analyzing the performance of the

Author statement

Iman Afshoon: Resources, Writing - Original Draft, Data curation, Writing- Original draft preparation, Software.

Mahmoud Miri: Conceptualization, Methodology, Investigation, Reviewing and Editing, Supervision, Project administration, funding acquisition.

Seyed Roohollah Mousavi: Resources, Writing - Original Draft, Data curation, Writing- Original draft preparation, Software.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References (81)

  • K. Duan et al.

    Boundary effect on concrete fracture and non-constant fracture energy distribution

    Eng. Fract. Mech.

    (2003)
  • R.A. Einsfeld et al.

    Fracture parameters for high-performance concrete

    Cement Concr. Res.

    (2006)
  • M. Ghasemi et al.

    Studying the fracture parameters and size effect of steel fiber-reinforced self-compacting concrete

    Construct. Build. Mater.

    (2019)
  • M.T. Kazemi et al.

    Fracture properties of steel fiber reinforced high strength concrete using work of fracture and size effect methods

    Construct. Build. Mater.

    (2017)
  • A. Carpinteri et al.

    Fracture behaviour of plain and fiber-reinforced concrete with different water content under mixed mode loading

    Mater. Des.

    (2010)
  • Z. Zhao et al.

    Effect of specimen size on fracture energy and softening curve of concrete: Part I. Experiments and fracture energy

    Cement Concr. Res.

    (2008)
  • B. Chen et al.

    Effect of aggregate on the fracture behaviour of high strength concrete

    Construct. Build. Mater.

    (2004)
  • J. Brown et al.

    Fracture toughness of cement paste and mortars

    Cement Concr. Res.

    (1973)
  • N. Isu et al.

    Influence of quartz particle size on the chemical and mechanical properties of autoclaved aerated concrete (II) fracture toughness, strength and micropore

    Cement Concr. Res.

    (1995)
  • T. Dittmer et al.

    The effect of coarse aggregate content and size on the age at cracking of bonded concrete overlays subjected to restrained deformation

    Construct. Build. Mater.

    (2014)
  • M. Elices et al.

    Effect of aggregate size on the fracture and mechanical properties of a simple concrete

    Eng. Fract. Mech.

    (2008)
  • M.H.A. Beygi et al.

    Evaluation of the effect of maximum aggregate size on fracture behavior of self-compacting concrete

    Construct. Build. Mater.

    (2014)
  • P.E. Petersson

    Fracture energy of concrete: practical performance and experimental results

    Cement Concr. Res.

    (1980)
  • J. Vahedi et al.

    An adaptive divergence-based method for structural reliability analysis via multiple Kriging models

    Appl. Math. Model.

    (2018)
  • Y. Moodi et al.

    Using Response Surface Methodology and providing a modified model using whale algorithm for estimating the compressive strength of columns confined with FRP sheets

    Construct. Build. Mater.

    (2018)
  • S. Khalilpour et al.

    A review on concrete fracture energy and effective parameters

    Cement Concr. Res.

    (2019)
  • R. Ince

    Prediction of fracture parameters of concrete by artificial neural networks

    Eng. Fract. Mech.

    (2004)
  • R. Ince

    Determination of concrete fracture parameters based on two-parameter and size effect models using split-tension cubes

    Eng. Fract. Mech.

    (2010)
  • Z.P. Bazant et al.

    Statistical prediction of fracture parameters of concrete and implications for choice of testing standard

    Cement Concr. Res.

    (2002)
  • Y.C. Zhou et al.

    A Bayesian Monte Carlo-based method for efficient computation of global sensitivity indices

    Mech. Syst. Signal Process.

    (2019)
  • E.I. Flores et al.

    A computational multi-scale approach for the stochastic mechanical response of foam filled honeycomb cores

    Compos. Struct.

    (2012)
  • L.G. Zhang et al.

    Efficient structural reliability analysis method based on advanced Kriging model

    Appl. Math. Model.

    (2015)
  • B. Echard et al.

    An active learning reliability method combining Kriging and Monte Carlo simulation

    Struct. Saf.

    (2011)
  • X. Huang et al.

    DSKmeans: a new kmeans-type approach to discriminative subspace Clustering

    Knowl. Base Syst.

    (2014)
  • S. Zahra et al.

    Novel centroid selection approaches for KMeans clustering based recommender systems

    Inf. Sci.

    (2015)
  • X. Wang et al.

    Effect of water–cement ratio, aggregate type, and curing temperature on the fracture energy of concrete

    Construct. Build. Mater.

    (2020)
  • A. Yan et al.

    Effect of fracture path on the fracture energy of high-strength concrete

    Cement Concr. Res.

    (2001)
  • M. Janssen et al.

    Fracture Mechanics VSSD

    (2006)
  • S. Xu et al.

    Crack extension resistance and fracture properties of quasi-brittle softening materials like concrete based on the complete process of fracture

    Int. J. Fract.

    (1998)
  • A.S. Rao et al.

    Fracture mechanics of fiber reinforced concrete: an overview

    International Journal of Engineering Innovations and Research

    (2014)
  • Cited by (23)

    • Generalized uncertainty in surrogate models for concrete strength prediction

      2023, Engineering Applications of Artificial Intelligence
    View all citing articles on Scopus
    View full text