Scholarly article on topic 'Prediction of Scaling Resistance of Concrete Modified with High-calcium Fly Ash Using Classification Methods'

Prediction of Scaling Resistance of Concrete Modified with High-calcium Fly Ash Using Classification Methods Academic research paper on "Materials engineering"

CC BY-NC-ND
0
0
Share paper
Academic journal
Procedia Computer Science
OECD Field of science
Keywords
{"High calcium fly ash" / "Freeze-thaw resistance" / "Scaling resistance" / "Machine learning" / Classification / "durability prediction"}

Abstract of research paper on Materials engineering, author of scientific article — Michał Marks, Maria Marks

Abstract The goal of the study was applying machine learning methods to create rules for prediction of the surface scaling resistance of concrete modified with high-calcium fly ash. To determine the scaling durability the Boras method, according to European Standard procedure (PKN- CEN/TS 12390-9:2007), was used. The results of numeral experiments were utilized as a training set to generate rules indicating the relation between material composition and the scaling resistance. The classifier generated by BFT algorithm from the WEKA workbench can be used as a tool for adequate classification of plain concretes and concretes modified with high-calcium fly ash as materials resistant or not resistant to the surface scaling.

Academic research paper on topic "Prediction of Scaling Resistance of Concrete Modified with High-calcium Fly Ash Using Classification Methods"

Procedia Computer Science

Volume 51, 2015, Pages 394-403

ICCS 2015 International Conference On Computational Science

Prediction of scaling resistance of concrete modified with high-calcium fly ash using classification methods

Michal Marks1 and Maria Marks2

1 Research and Academic Computer Network, Wawozowa 18, 02-796 Warsaw, Poland

mmarks@nask.pl

2 Institute of Fundamental Technological Research, Polish Academy of Sciences, Pawinskiego 5B,

02-106 Warsaw, Poland mmarks@ippt.gov.pl

Abstract

The goal of the study was applying machine learning methods to create rules for prediction of the surface scaling resistance of concrete modified with high-calcium fly ash. To determine the scaling durability the Boras method, according to European Standard procedure (PKN-CEN/TS 12390-9:2007), was used. The results of numeral experiments were utilized as a training set to generate rules indicating the relation between material composition and the scaling resistance. The classifier generated by BFT algorithm from the WEKA workbench can be used as a tool for adequate classification of plain concretes and concretes modified with high-calcium fly ash as materials resistant or not resistant to the surface scaling.

Keywords: high calcium fly ash, freeze-thaw resistance, scaling resistance, machine learning, classification, durability prediction

1 Introduction

The application of high calcium fly ash (HCFA) for partial replacement of Portland cement in concrete could provide a number of environmental benefits (reduced CO2 emissions during cement production, reduced consumption of cement clinker, saving natural resources). Moreover the high calcium fly ash is produced as by-product of power generation in brown coal burning plants. On the other hand, this type of ash is usually characterized by low silica content and an increased content of sulfur compounds. It could be used in concrete following the requirements of ASTM C618 class C, but it does not meet the requirements defined in standard EN 450-1 and is not commonly used in European countries despite positive examples of its suitability provided by some researchers. It was shown [10] that in the case of cement replacement with HCFA, the compressive strength of concrete was increased if the content of active silica in the fly ash was higher than that in the cement. Similar results were obtained earlier by Naik et al. [9]: partial replacement of cement by fine grained HCFA resulted in the same or better compressive

394 Selection and peer-review under responsibility of the Scientific Programme Committee of ICCS 2015

© The Authors. Published by Elsevier B.V.

doi:10.1016/j.procs.2015.05.259

strength of concrete; the results for drying shrinkage were also positive. The optimization of fineness coupled with the adjustment of water content were found as the key parameters of effective utilization of high-calcium fly ashes for strength maximization of cement mortars [2].

There are only few papers on the freeze-thaw resistance of concrete containing calcareous fly ash. According to Yazici [17], calcareous fly ash added at a rate of between 30% and 60% improves the resistance of non-air-entrained concrete to internal damage due to freeze-thaw cycling. This, however, is contrary to the findings described in [4], namely that the addition of calcium-type fly ash has an adverse effect on the freeze-thaw resistance of non-air-entrained concrete [1]. All mentioned results concentrate mostly on internal frost resistance not on the surface scaling resistance, so providing rules for surface scaling resistance assessment is even more significant.

Prediction of engineering properties of composite materials is usually based on experimental test results with a reference to observed material microstructure. The relevant material characteristics can be extracted from experimental dataset using various computational intelligence methods, developed for the last two decades for various engineering applications [5, 8]. Artificial neural networks were successfully used for prediction of compressive strength of concrete containing coal ash [12] or silica fume [14]. Using neural networks and optimization technologies it was possible to search for the optimum mixture of concrete - the mixture with the lowest cost and required performance, such as strength and slump [18]. Machine learning methods were also tested on classification of concrete modified by fluidized bed fly ash as materials of adequate resistance to chloride penetration [6] and resistance to the surface scaling [7].

The aim of this research was to generate rules using machine learning algorithms to evaluate the de-icing salt scaling resistance of concrete manufactured using calcareous fly ash as a mix additive or introduced as one of the main constituents of blended cements.

2 Composition of concrete mixes and test results of scaling resistance coefficient

The surface scaling resistance in concrete specimens with different content of high calcium fly ash was experimentally measured. Concrete mixes were prepared with replacement of 15% or 30% of cement mass by HCFA. Experimental tests were performed on several mixes, based on two types of Portland cement C1 - CEM I 42.5R (with 10 percent of C3A content) or C2 -CEM I 42.5 HSR NA (with 2 percent of C3A content), siliceous sand fraction 0 + 2 mm and the amphibolite as a coarse aggregate (two fractions 2 + 8 mm and 8 +16 mm). Because of expected variability of ash properties six lots of high calcium fly ash were tested from different deliveries from the Belchatow power plant, namely S1 - March 2010, S2 - May 2010, S3 - June 2010, S4 -November 2010, S5 - March 2011 and S6 - September 2011. HCFA was used as an additive to concrete mix in unprocessed form (as collected) and after grinding (physical properties of ash before and after grinding are given in Table 1). Chemical properties of HCFA lots are shown in Table 2.

In order to determine the scaling resistance of concrete specimens a series of tests were performed according to European Standard procedure (PKN-CEN/TS 12390-9:2007). The upper horizontal surface of the specimens (the cut surface) was exposed to freezing and thawing while the remaining surfaces were isolated against humidity and heat transfer. After 28 days of curing the top exposed surface was covered with 3% NaCl solution. The mean mass of scaled material after 28 (m28) and 56 (m56) cycles is used for evaluating the scaling resistance, according to the criteria presented in Table 3. The Boras method, defined by the Swedish

Table 1: Physical properties of high-calcium fly ashes before and after processing, [3]

Batch Fly ash designation Density Fineness - the residue Specific surface

[g/cm3] on sieve 45 ¡m [%] by Blaine [cm2/g]

S1 : unprocessed 2.62 38.0 2860

S1 S1io: grinded 10 min 2.77 23.0 3500

S128: grinded 28 min 2.75 10.5 3870

S2 S2: unprocessed 2.58 35.4 4400

S215: grinded 15 min 2.70 13.3 6510

S3 S3: unprocessed 2.64 55.6 1900

S320: grinded 20 min 2.71 20.0 4060

S4 S4: unprocessed 2.60 57.2 1900

S420: grinded 20 min 2.63 16.7 4700

S5 S5: unprocessed 2.60 46.3 2370

S51b: grinded 15 min 2.67 20.8 3520

S6 S6: unprocessed 2.41 59.2 2190

S623: grinded 23 min 2.50 20.3 4000

standard (SS 137244 ,1995) indicates 4 concrete categories dependent of the amount of scaled material collected and weighed after given numbers of freeze/thaw cycles.

In Table 4 database containing data on composition of the concrete mixes and mass of scaled material after 28 and 56 cycles is presented. The estimation of the surface scaling resistance, based on Boras criterion presented in Table 3, is placed in the last column of Table 4. The database presented in Table 4 is a general database, which can be transformed into a "working database" by attributes transformation and selection.

Table 2: The chemical composition of high-calcium fly ashes determined using XRF method. Fly ash sampling date and bath designation, [3]

Fly ash sampling date and batch designation

Component 16.03.2010 19.05.2010 28.06.2010 10.11.2010 25.03.2011 13.09.2011

S1 S2 S3 S4 S5 S6

LOI 2.56% 3.43% 1.85% 2.67% 2.12% 2.73%

SiO2 33.62% 35.41% 40.17% 45.17% 40.88% 47.2%

AI2O3 19.27% 21.86% 24.02% 20.79% 19.00% 20.54%

Fe 2O3 5.39% 6.11% 5.93% 4.58% 4.25% 4.47%

CaO 31.32% 25.58% 22.37% 20.60% 25.97% 19.11%

MgO 1.85% 1.49% 1.27% 1.49% 1.73% 1.51%

SO3 4.50% 4.22% 3.07% 2.96% 3.94% 2.26%

K2O 0.11% 0.13% 0.20% 0.19% 0.14% 0.15%

Na2O 0.31% 0.16% 0.15% 0.23% 0.13% 0.12%

P2O5 0.17% 0.16% 0.33% 0.14% 0.10% 0.12%

TiO2 1.21% 1.22% 1.01% 1.37% 1.52% 1.43%

Mn2O3 0.07% 0.06% 0.06% 0.06% 0.04% 0.03%

SrO 0.20% 0.17% 0.16% 0.13% 0.17% 0.10%

ZnO 0.02% 0.02% 0.02% 0.02% 0.01% 0.02%

CaO free 2.87% 1.24% 1.46% 1.18% 1.07% 1.00%

Table 3: Criteria of the surface scaling resistance evaluation

Scaling resistance Requirements

Very good m56 < 0.10 kg/m2

Good m56 < 0.20 kg/m2 or m56 < 0.50 kg/m2 and m56/m28 < 2

Acceptable m56 < 1.00 kg/m2 and m56/m28 < 2

Unacceptable m.56 > 1.00 kg/m2 and m^/m28 > 2

3 Prediction of engineering properties of composite materials based on Machine Learning techniques

3.1 Introduction to machine learning

Determining relationship between material composition and the surface scaling resistance of concrete is a difficult and time-consuming process even in case a small dataset as presented in Table 4. This task can be done manually, however using an automatic data exploration tools is much more efficient. A branch of artificial intelligence concerned with the applying algorithms that allow computers to evolve patterns based on empirical data is called machine learning.

A major focus of machine learning research is to automatically learn to recognize complex patterns and make intelligent decisions based on dataset, i.e. collection of logically related records. Each record can be called an example or instance and each one is characterized by the values of predetermined attributes. The major difficulty is the fact that the set of all possible behaviors given all possible inputs is too large to be covered by the set of observed examples (training data). Hence the learner must generalize from the given examples, so as to be able to produce a useful output in new cases.

Patterns recognition associated usually with classification is the most popular example of utilizing machine learning. However machine learning or, more general, statistical algorithms can support the knowledge discovery on different stages from outliers detection and attributes (features) selection to knowledge modeling and models validation.

3.2 Attributes selection

Attribute selection, also known as feature selection, variable selection or feature reduction, is the technique of selecting a subset of relevant features for building robust learning models. By removing most irrelevant and redundant attributes from the data, feature selection helps improve the performance of learning models by: speeding up learning process and alleviating the effect of the curse of dimensionality. Moreover the irrerelevant attributes degrade the performance of state-of-the-art decision tree and rule learners [15].

3.3 Classification

As noted earlier in section 3.1 classification is the most common type of machine learning applications. The aim of the classification process is to learn a way of classifying unseen examples based on the knowledge extracted from the provided set of classified examples. In order to extract the knowledge from the provided dataset the attribute set characterizing the example has to be divided into two groups: the class attribute and the non-class attributes. For an unseen examples only non-class attributes are known, therefore the aim of data mining algorithms is

Table 4: The database of composition of concrete mixes and properties of hardened concretes

Concrete Content [kg/m3] Mass of scaled material Scaling resistance

Cement HCFA Aggregate Water Plasticizer Air-entraining after 28 cycles after 56 cycles

mix C HCFA K W P A m28 m56 resistance

A-0 350 C1 no 1890 158 2,1 0,49 0,012 0,016 very good

A-15 298 C1 133 S1 1800 158 2,682 0,9238 0,035 0,114 good

B-15 298 C1 133 S1io 1800 158 2,682 2,086 0,123 0,171 good

C-15 298 C1 133 S128 1800 158 2,086 3,874 0,164 0,246 good

A-30 245 C1 263 S1 1710 158 5,145 1,96 0,098 0,424 unacceptable

B-30 245 C1 263 S1io 1710 158 4,165 2,94 0,095 1,173 unacceptable

C-30 245 C1 263 S128 1710 158 3,43 4,165 0,347 1,234 unacceptable

D-15 298 C1 133 S2 1800 158 8,642 3,874 0,016 0,333 unacceptable

E-15 298 C1 133 S2i5 1800 158 5,066 6,258 0,522 1,035 unacceptable

SE-0 340 C1 no 1825 153 0,34 0,204 0,054 0,071 very good

SE-30S 238 C1 102 S2 1808 153 2,142 0,6902 0,68 1,306 unacceptable

H-0 350 C2 no 1880 175 0 0,7 0,38 0,46 good

H-15M 298 C2 75 S320 1847 175 0 4,172 0,16 0,19 good

H-15S 298 C2 75 S3 1847 175 0 1,49 1,86 2,65 unacceptable

H-30M 245 C2 150 S320 1813 175 0,49 4,165 2,03 2,56 unacceptable

H-30S 245 C2 150 S3 1813 175 0,49 1,225 2,31 3,2 unacceptable

N-0 350 C1 no 1880 175 0,35 0,35 0,41 0,54 acceptable

N-15M 298 C1 75 S320 1847 175 0,298 2,98 0,15 0,16 good

N-15S 298 C1 75 S3 1847 175 0,298 0,894 0,41 0,5 acceptable

N-30M 245 C1 150 S320 1813 175 0,98 4,165 0,28 0,59 unacceptable

N-30S 245 C1 150 S3 1813 175 1,47 1,225 0,07 0,17 good

K-0 350 C1 no 1880 175 0,35 0,35 0,405 0,536 acceptable

K-15M 298 C1 75 S420 1847 175 0,298 4,172 0,695 0,727 acceptable

K-15S 298 C1 75 S4 1847 175 0,596 1,49 2,328 3,047 unacceptable

K-30M 245 C1 150 S420 1813 175 0,735 6,125 1,391 1,611 unacceptable

K-30S 245 C1 150 S4 1813 175 0,735 2,205 1,705 2,14 unacceptable

SEP-0 340 C1 no 1825 153 0,34 0,068 0,315 0,46 good

SEP-V-30S 238 C1 102 S5 1808 153 1,19 0,1428 1,537 1,888 unacceptable

SEP-V-30M 238 C1 102 S5i5 1808 153 0,238 0,7854 0,712 0,86 acceptable

SEP-VI-30S 238 C1 102 S6 1808 153 3,094 0,4998 0,45 0,505 acceptable

SEP-VI-30M 238 C1 102 S623 1808 153 1,19 0,6426 0,549 0,653 acceptable

to build such a knowledge model that allows predicting the example class membership based only on non-class attributes.

The knowledge model is dependent on the way how the classifier is constructed and it can be represented by decision trees (e.g. algorithm C4.5, [11]), classification rules (the algorithm AQ21 [16]) or many other representations. Regardless of the representation both decision trees and classification rules algorithms create hypotheses.

In considered problem the surface scaling resistance of concrete (class attribute) depending on material composition and additives properties (non-class attributes) is searched. This search can be done using one of many software suites available on the market such as: Weka, RapidMiner, LIONsolver, KNIME, See5 and many others. In our experiments we decided to utilize WEKA workbench [15]. This software suite contains almost one hundred algorithms representing different approaches to machine learning such as: bayesian networks, decision trees or association rules. In our experiments we concentrated on the representative of decision trees classifier - the BFT algorithm [13] - the well known modification of the last publicly available version of a C4.5 method developed by J. Ross Quinlan [11]. This algorithm was compared with other available in Weka algorithms in section 4.2.

4 Searching for the rules describing the surface scaling resistance of concrete modified with high-calcium fly ash

4.1 Attributes selection

In Table 4 the dataset with 6 composition attributes is presented. However these attributes are expressed in absolute values, so they are not appropriate for comparisons. To alleviate this inconvenience the modified database with concrete mix composition expressed in percentage, additions to cement ratio (A2C) and additions properties like specific surface by Blaine (BL) and density (DE) is provided - (Table 5). Due to small number of instances we decided to express class attribute in the working database in a binary form, i.e. all examples with class very good, good and acceptable are represented as resistant instances and all example with unacceptable resistance as not resistant.

It is clear that for database with a few dozens of instances this number of attributes (9 attributes) is too large. Some attributes can be eliminated but it is important to eliminate the most irrelevant attributes. So that we decided to evaluate a subset of attributes using best first and exhaustive approaches to features selection. Best first method searches the space of attributes by greedy hillclimbing augmented with backtracking facility. In both cases the CfsSubsetEvaluator, provided by Weka, was used to assess the predictive ability of each attribute individually and the degree of redundancy among them, preferring sets of attributes that are highly correlated with the class but have low intercorrelation. Both methods of searching (best first and exhaustive) resulted in selection of C, HCFA, A2C, BL and DE as the most relevant attributes.

Therefore, in order to generate rules describing the surface scaling resistance of concrete modified with high-calcium fly ash the subset of attributes (C, HCFA, A2C, BL, DE, class) from the database (marked with bold in Table 5) is used. The shrunken database contains 31 records, each one described by 5 numerical and one nominal attributes. The last attribute - class -denotes a class and can take one of two values (resistant, not resistant).

Table 5: The database of composition of concrete mixes and properties of hardened concretes

mix C HCFA K W P A A2C BL DE class

A-0 14.58 0 78.73 6.58 0.09 0.02 0 0 0 resistant

A-15 12.46 5.56 75.23 6.6 0.11 0.04 30.9 2860 2.62 resistant

B-15 12.45 5.56 75.2 6.6 0.11 0.09 30.9 3500 2.77 resistant

C-15 12.44 5.55 75.16 6.6 0.09 0.16 30.9 3870 2.75 resistant

A-30 10.28 11.04 71.76 6.63 0.22 0.08 51.8 2860 2.62 not resistant

B-30 10.28 11.04 71.76 6.63 0.17 0.12 51.8 3500 2.77 not resistant

C-30 10.28 11.03 71.74 6.63 0.14 0.17 51.8 3870 2.75 not resistant

D-15 12.41 5.54 74.95 6.58 0.36 0.16 30.9 4400 2.58 not resistant

E-15 12.41 5.54 74.99 6.58 0.21 0.26 30.9 6510 2.70 not resistant

SE-0 14.66 0 78.71 6.6 0.01 0.01 0 0 0 resistant

SE-30S 10.33 4.43 78.48 6.64 0.09 0.03 30 4400 2.58 not resistant

H-0 14.55 0 78.15 7.27 0 0.03 0 0 0 resistant

H-15M 12.42 3.13 76.98 7.29 0 0.17 20.1 4060 2.71 resistant

H-15S 12.43 3.13 77.07 7.3 0 0.06 20.1 1900 2.64 not resistant

H-30M 10.26 6.28 75.93 7.33 0.02 0.17 38 4060 2.71 not resistant

H-30S 10.27 6.29 76.03 7.34 0.02 0.05 38 1900 2.64 not resistant

N-0 14.55 0 78.15 7.27 0.01 0.01 0 0 0 resistant

N-15M 12.43 3.13 77.01 7.3 0.01 0.12 20.1 4060 2.71 resistant

N-15S 12.44 3.13 77.08 7.3 0.01 0.04 20.1 1900 2.64 resistant

N-30M 10.26 6.28 75.92 7.33 0.04 0.17 38 4060 2.71 not resistant

N-30S 10.27 6.29 75.99 7.34 0.06 0.05 38 1900 2.64 resistant

K-0 14.55 0 78.15 7.27 0.01 0.01 0 0 0 resistant

K-15M 12.42 3.13 76.98 7.29 0.01 0.17 20.1 4700 2.63 resistant

K-15S 12.43 3.13 77.05 7.3 0.02 0.06 20.1 1900 2.60 not resistant

K-30M 10.25 6.28 75.86 7.32 0.03 0.26 38 4700 2.63 not resistant

K-30S 10.27 6.29 75.99 7.33 0.03 0.09 38 1900 2.60 not resistant

SEP-0 14.67 0 78.72 6.6 0.01 0 0 0 0 resistant

SEP-V-30S 10.34 4.43 78.53 6.65 0.05 0.01 30 2370 2.6 not resistant

SEP-V-30M 10.34 4.43 78.54 6.65 0.01 0.03 30 3520 2.67 resistant

SEP-VI-30S 10.33 4.43 78.45 6.64 0.13 0.02 30 2190 2.41 resistant

SEP-VI-30M 10.34 4.43 78.51 6.64 0.05 0.03 30 4000 2.50 resistant

4.2 Classification

As it was mentioned in section 3.3 surface scaling resistance of concrete depending on material composition can be searched using one of many software suites available on the market and we decided to utilize WEKA workbench. The WEKA workbench provides a collection almost one hundred algorithms supporting classification. They belong to different types like: bayesian classifiers, rules classifiers, trees classifiers or meta classifiers. In our research we decided to determine the surface scaling durability of concrete using selected 8 algorithms belonging to three different types of algorithms. As an training set all the instances from the database (Table 5) were considered. The classification accuracy was evaluated using leave-one-out cross validation [15]. The obtained results are collected in Table 6.

Table 6: Results obtained for different classifiers from Weka Workbench

Bayesian classifiers Trees classifiers Rules classifiers

Classifier BayesNet NaiveBayes BFTree FT LADTree J48 JRip PART

Accuracy 61.29 61.29 74.19 58.06 74.19 70.97 64.52 67.74

The best accuracy equal almost 75% was obtained using BFT and LADTree algorithms. The decision tree generated by BFT algorithm is presented in Fig. 1.

<=30.45 >30.45

not resistant (10/1)

<=2045 >2045

Figure 1: The decision tree for scaling resistance generated by BFT

In Fig. 1 the first number in brackets denotes the number of examples set covered by a selected leaf, and the second number - just after the sign " number of incorrectly classified instances (negative examples).

The obtained decision tree can be easily transformed into the following rules: [class = resistant] rule 1 [C> 12.43%]: p =10, n = 0,

rule 2 [C<12.43%] and [A2C< 30.45%] and [BL>2045] : p = 4, n = 2. [class = not resistant]

rule 1 [C<12.43%] and [A2C>30.45%] : p = 9, n = 1,

rule 2 [C<12.43%] and [A2C<30.45%] and [BL<2045] : p = 2, n = 0.

where p denotes the number of positive examples covered by the rule (i.e. the number of records from this class satisfying the rule), n denotes the number of negative examples covered by the rule (i.e. the number of records from the other classes satisfying the rule).

The obtained decision rules determines conditions concretes have to fulfill to provide appropriate surface scaling resistance.

The class of resistant characterize:

• concretes with percent of cement in mix composition greater then 12.43% (C> 12.43%),

algorithm

from the training /" - indicates the

among other things - all plain concretes (with no cement mass replaced with HCFA),

• concretes with percent of cement in mix composition lower then 12.43% (C< 12.43%) where not more then 30.45% of cement mass (A2C < 30.45) was replaced by high-calcium fly ash where specific surface according to Blaine is greater then 2045 cm2/g (BL > 2045).

The class of not resistant characterize:

• concretes with percent of cement in mix composition lower then 12.43% (C< 12.43%) where more then 30.45% of cement mass (A2C < 30.45) was replaced by HCFA,

• concretes with percent of cement in mix composition lower then 12.43% (C<12.43%) where not more then 30.45% of cement mass (A2C < 30.45) was replaced by high-calcium fly ash where specific surface according to Blaine is lower then 2045 cm2/g (BL < 2045).

Using the method leave-one-out (n = 31) we obtained the classification accuracy equal 74.19%. The result obtained from a test set is often displayed as a two-dimensional confusion 'matrix with a row and a column for each class. Each matrix element shows the number of test examples for which the actual class is the row and the predicted class is the column. The sum of the numbers down the main diagonal divided by the total number of test examples determine classification accuracy. The confusion matrix of the solved problem is determined in the following form:

resistant not resistant

resistant 12 5

not resistant 3 11

Such a result can be considered satisfactory in respect to the very limited number of records in the database.

5 Conclusions

The rules generated by algorithm BFT from the WEKA workbench provided means for adequate classification of plain concretes and concretes modified with high-calcium fly ash as materials resistant or not resistant to surface scaling. According to generated rules may state that obtaining concrete mixes with good surface scaling resistance require using not more then 30% of HCFA. Moreover HCFA additions should be characterized by not too small specific surface. In majority of cases appropriate specific surface can be obtained by additions grinding. Due to a small number of tested specimens the rules are applicable only to concrete mix compositions of similar binder content. Further tests are needed in order to enlarge the experimental database and to cover broader range of concrete compositions, especially that plain concretes with unacceptable surface scaling resistance was not represented in considered dataset.

Acknowledgements

Authors would like to express their gratitude to IFTR Polish Academy of Sciences team for providing access to their experimental dataset collected during realization of the research project "Innovative cement based materials and concrete with high calcium fly ashes" co-financed by the European Union from the European Regional Development Fund.

References

[1] M. Dabrowski and Michal A. Glinicki. Air void system parameters and frost resistance of air-entrained concrete containing calcareous fly ash. Roads and Bridges - Drogi i Mosty, Vol. 12, no. 1:41-55, 2013.

[2] Burak Felekoglu, Selcuk Turkel, and Hasan Kalyoncu. Optimization of fineness to maximize the strength activity of high-calcium ground fly ash - portland cement composites. Construction and, Building Materials, 23(5):2053-2061, 2009.

[3] K. Gibas and M. A. Glinicki. Influence of high-calcium fly ashes on the chloride ion penetration into concrete. In Proceedings of the Int. Conf. Non-Traditional Cement & Concrete IV, pages 419-428, Brno, 2011.

[4] Z. Giergiczny, K. Synowiec, and A. ak. Suitability evaluation of calcareous fly ash as an active mineral additive to concrete. Roads and Bridges - Drogi i Mosty, Vol. 12, no. 1:83-97, 2013.

[5] Lawrence J. Kaetzel and James R. Clifton. Expert/knowledge based systems for materials in the construction industry: State-of-the-art report. Materials and Structures, 28(3):160-174, 1995.

[6] M. Marks, D. Jozwiak-Niedzwiedzka, and M. A. Glinicki. Automatic categorization of chloride migration into concrete modified with cfbc ash. Computers and Concrete, 9(5):375-387, 2012.

[7] M. Marks, D. Jozwiak-Niedzwiedzka, M. A. Glinicki, J. Olek, and M. Marks. Assessment of scaling durability of concrete with cfbc ash by automatic classification rules. Journal of Materials in Civil Engineering, 24(7):860-867, July 2012.

[8] Ralf Mikut and Markus Reischl. Data mining tools. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 1(5):431-443, 2011.

[9] T. R. Naik, S. S. Singh, and M. M. Hossain. Properties of high performance concrete systems incorporating large amounts of high-lime fly ash. Construction and Building Materials, 9(4):195-204, 1995.

[10] V. G. Papadakis. Effect of fly ash on portland cement systems - part II. high-calcium fly ash. Cement and Concrete Research, 30(10):1647-1654, 2000.

[11] J. R. Quinlan. C4.5: Programs for Machine Learning. Morgan Kaufmann Series in Machine Learning. Morgan Kaufmann Publishers, 1993.

[12] Marta Sebastia, Inaki Fernandez Olmo, and Angel Irabien. Neural network prediction of un-confined compressive strength of coal fly ash-cement mixtures. Cement and Concrete Research, 33(8):1137-1146, 2003.

[13] Haijian Shi. Best-first decision tree learning. Technical report, University of Waikato, 2007.

[14] Ilker Bekir Topcu and Mustafa Saridemir. Prediction of mechanical properties of recycled aggregate concretes containing silica fume using artificial neural networks and fuzzy logic. Computational Materials Science, 42(1):74-82, 2008.

[15] I. H. Witten and E. Frank. Data Mining: Practical Machine Learning Tools and Techniques. Morgan Kaufmann Series in Data Management Systems. Morgan Kaufman, 2005.

[16] J. Wojtusiak. Aq21 user's guide. reports of the machine learning and inference laboratory. Technical Report MLI 04-3, George Mason University, 2004.

[17] Halit Yazici. The effect of silica fume and high-volume class c fly ash on mechanical properties, chloride penetration and freezethaw resistance of self-compacting concrete. Construction and Building Materials, 22(4):456 - 462, 2008.

[18] I-Cheng Yeh. Computer-aided design for optimum concrete mixtures. Cement and Concrete Composites, 29(3):193 - 202, 2007.