The Use of Artificial Neural Network for Predicting the Thermal Conductivity of Cement Based Mortar with Natural Zeolites

Zeolites, in their natural state, have been used in construction materials since ancient times. The pozzolanic activity of zeolites and their use as supplementary cementitious materials has been investigated over the past three decades. The indoor comfort provided by a modern life style comes at a cost of consuming enormous amounts of energy. Materials with improved insulation properties are continuously researched and the use of zeolites showing encouraging results. The paper presents experimental results obtained on different cement-based mortar mixes incorporating natural zeolites. The main parameters of the research were: substitution of cement and sand by natural zeolites, three different replacement percentages for cement and sand (10%, 20% and 30%) and the curing age of mortar specimens (14 days, 21 days and 28 days). An artificial neural network (ANN) was developed to predict the values of the thermal conductivity by taking into account parameters such as density, humidity and surface temperature of the mortar samples. The ANN was able to accurately predict the experimental results for the thermal conductivity of cement-based mortars with natural zeolites.


Introduction
Zeolites, in their natural state, have been used in construction materials since ancient times.Zeolites are hydrated alumino-silicates and, therefore, contain large amounts of reactive SiO2 and Al2O3.These high concentrations give them pozzolanic properties similar to fly-ash, silica fume and metakaolin [1].The pozzolanic activity of zeolites and their use as supplementary cementitious materials has been investigated over the past three decades with results in terms of durability and mechanical properties of cement-based materials varying over a wide range mainly due to zeolite source and replacement percentages [2,3].Zeolites showed promising results in mitigating the autogenous shrinkage of highperformance concrete due to their ability to retain water and gradually release it in time [4,5] which helps to significantly reduce the shrinkage induced cracking.Moreover, recent studies showed the possibility of using zeolites in self-healing cement-based mortars, acting as a bacterial carrier for the self-healing process [6].
The indoor comfort provided by a modern life style comes at a cost of consuming enormous amounts of energy which, inevitably, leads to large CO2 emissions.Although, recently, an increasing quantity of required energy comes from renewable sources, there is still a stringent need to develop new materials [7] with improved thermal insulation properties and to correctly assess their performance.Even in this aspect, zeolites showed promising results and research is still on-going [8].
The field of computational intelligence expanded significantly during the last decade, with leaps in the development of artificial intelligence methods.These methods gradually found their applications in a variety of scientific fields, civil engineering being one of them.Various mechanical, durability, cleaning properties of construction materials are now able to be accurately predicted by means of evolutionary approaches (neural networks) as a subset of computational intelligence.As such, artificial neural networks (ANN) are used in a wide variety of civil engineering problems, mainly related to concrete properties considering the highly heterogeneity of the material [9,10].The development of a good ANN architecture could be beneficial in terms of long-term prediction of sought material parameters and, therefore, significantly reduce the waiting time.Of course, some long-term material characteristics could be determined by means of accelerated tests, e.g.durability issues, but those too require specialized equipment and a correct set-up of the experiment.
The paper presents experimental results obtained on different cement-based mortar mixes incorporating natural zeolites either as a substitute for cement or as a substitute for sand.As such, the main parameters of the research were: substitution of cement and sand by natural zeolites, three different replacement percentages for cement and sand (10%, 20% and 30%) and the curing age of mortar specimens (14 days, 21 days and 28 days).The thermal conductivity of all mixes was measured on specially cast cylindrical specimens.An ANN was developed, trained and used to predict the values of the thermal conductivity by taking into account parameters such as density, humidity and surface temperature of the mortar samples.The ANN was able to accurately predict the experimental results for the thermal conductivity of cement-based mortars with natural zeolites.

Mix proportions
The mortar mix consisted in a 1:4:0.6 mix, by volume, of constituent parts: CEM II/B-M 42.5R (rapid hardening composite cement), natural sand with a maximum particle size diameter of 4 mm and water.An initial water/binder ratio of 0.6 was considered but, depending on the consistency of the mortar, it was adjusted to obtain a mortar that could be easily cast in 70 × 30 mm (diameter × thickness/height) cylindrical steel molds.This was deemed necessary especially for mixes where cement was replaced by natural zeolites.
The natural zeolite used in this study was based on Clinoptilolite.It had a mean particle distribution of 29 μm, when used to replace cement, and a maximum particle size of 4 mm, when used as sand replacement.The mix proportions, expressed as mass quantities, are presented in [11], for cement replacement, and in [12], for sand replacement.

Methods
A total of nine specimens were cast for each mix proportion, resulting in a total of 63 cylindrical specimens.As previously stated, the main parameters of the research were: cement or sand replacement by natural zeolites, replacement percentage and curing age.There were seven mix proportions considered: a control mix, without any zeolites, three mix proportions where cement was replaced by micronized zeolite (10%, 20% and 30% replacement, by volume) and three mix proportions where sand was replaced by natural zeolites (10%, 20% and 30% replacement, by volume).The mortar specimens were demolded after 24 hours from casting and placed in tap water for curing up to 28 days.
The thermal conductivity coefficient was measures by means of an ISOMET 2114 equipment (Figure 1) at the ages of 14, 21 and 28 days for all 7 mix proportions.Six individual measurements/readings were taken for each mortar sample, for a total of 1134 data entries.The cylindrical specimens were taken out of water, wiped by a cloth to remove the excess surface water and kept at room temperature for 12 hours before the thermal conductivity coefficient was measured.This would ensure that a surface dry condition of the specimens was obtained, without influencing the hydration process of cement inside the specimen, so that the obtained results would not be biased by the presence of water.
The specimens were measured, using a digital calliper with a precision of 0.1 mm, and weighted, with a precision of 0.1 grams, before each measurement of the thermal coefficient.This would help obtain of the bulk density of each specimen.The surface moisture content and temperature of the mortar sample were measured by means of a Merlin EVO CC equipment (Figure 2).Density, moisture content and temperature are influencing parameters for the thermal coefficient.

3
Using the Matlab and the Machine Learning Tool, the experimental data were used for creating a neural network (NN) feed-forward architecture.The data, 1134 values, was divided into 70% for training (851 values), 15% for testing (141 values) and 15% for validation (142 values).Several algorithms and configurations were considered for the learning process of the NN but the best solution found was a network with 12 neurons on the hidden layer and trained with the Levenberg-Marquardt algorithm.After learning the performance of the network were: Mean Squared Error MSE = 0.02 and Mu=1e-05.The schematic layout of the NN is presented in Figure 3.The MSE method has been frequently used and generally accepted as a method for assessing the performance of ANNs in different application fields [13,14].The MSE is a metric that tells the distance, or how far apart, are the predicted values from the observed, in this case experimentally determined, values in a dataset.According to earlier studies the number of neurons in the hidden layer is an essential parameter influencing the accuracy of the ANN [15].A too large number of neurons in the hidden layer may lead to overfitting while a too low number may result in underfitting issues [16].

Bulk density
The bulk density of all considered mixes at different curing ages are presented in Figure 4.The values were obtained by averaging the measurements from all 9 samples belonging to each mix proportion.In Figure 4 ZP stands for Zeolite Powder, used to replace the cement and having a mean particle size of 29 μm, and ZA stands for Zeolite Aggregate.The numbers following each notation represents the replacement percentage, by volume.
It can be observed that increasing the replacement percentage, of either cement or aggregate, resulted in a decrease in the value of bulk density.This decrease was more significant in case of replacing sand (ZA series).None of the considered mix proportions exhibited large significant variation in the values of bulk density from one curing age to another.Similar trends were previously reported [12].

Surface moisture
Figure 5 presents the surface moisture of each mix proportion as an average of nine individual measurements.It can be seen that ZA series exhibits the highest values of the surface moisture.Taking into account the porous structure of the zeolite aggregates and the fact that is absorbs the water during the mixing process to gradually release it over time, the obtained results are in line with previously reported data [17].The ZP series exhibited the lowest values of surface moisture.The higher the replacement percentage of cement by natural micronized zeolite, the lower the surface moisture.This trend is opposite to the one recorded for the ZA series.

Artificial neural network (ANN)
In order to correctly design the structure of the ANN, a statistical analysis was run on the influencing parameters upon the values of the thermal conductivity.According to the generally agreed knowledge, the thermal conductivity of a material depends on its density, moisture content and temperature.Taking into account that all specimens were kept in identical laboratory conditions, both during curing and measuring procedures, the effect of temperature could be neglected at this stage of the research.Hence, a regression analysis was conducted to assess the influence of bulk density and surface moisture on the obtained values for the thermal conductivity.Both parameters are highly dependent on the zeolite content of the mix proportion.
Figure 6a presents the spread of the obtained values of the thermal conductivity as function of bulk density for all 63 specimens.A sixth-degree polynomial function was used for regression analysis resulting in a value of R 2 = 0.677.A similar plot is presented in Figure 6b in terms of surface moisture.The obtained value of R 2 was 0.094.
It was, therefore, concluded that the moisture content did not influence the measured values of the thermal conductivity.The moisture content would affect the internal structure of the material in the long term but it did not show significant influence at the age of 28 days.This conclusion was also supported by the fact that all specimens were kept in water, therefore they were subjected to the same curing conditions.Therefore, the ANN configuration presented in Figure 3 was considered.The best validation performance of the ANN is presented in Figure 7, whereas the obtained error histogram is shown in Figure 8. From the two graphs it could be concluded that the error spread was limited, indicating a good accuracy of the results provided by the ANN.A similar trend was observed during the testing and validation stages of the ANN by using the assigned sets of data.

Conclusions
The paper presents experimental results in terms of thermal conductivity obtained on different cementbased mortar mixes incorporating natural zeolites either as a substitute for cement or as a substitute for sand and the use of ANN to predict the obtained values.The parameters of the research work are: the replacement percentage, the curing age and the component of the mortar to be replaced by the natural zeolites.
Replacing cement by micronized zeolite results in bulk density values similar to the reference mix.At the same time, the moisture content is smaller owing to a denser structure of the obtained mortar.When sand is replaced by natural zeolite the bulk density decreases with the increase in the replacement percentage but the moisture content increases.This can be explained by the porous structure of the zeolite which leads to the formation of void inside the mortar and facilitates the absorption of water.
The developed ANN is able to accurately predict the experimental data in terms of values of the thermal conductivity.Based on regression analysis it is concluded that the bulk density plays a significant role in the outcome of the results whereas moisture content and temperature are less significant.Taking into account the fact that all specimens are subjected to identical curing conditions, the moisture content and temperature are expected to have smaller influence on the thermal conductivity.During all stages of learning (training, validation and test), the data obtained from the ANN shows a good fit to the experimental results.The correlation coefficient has a value of approximately 0.95.This means that the proposed structure and algorithm can be further expanded and used to predict the values of the thermal conductivity at later ages.

Figure 4 .
Figure 4. Variation of bulk density considering curing time and mix proportion

Figure 5 .
Figure 5. Variation of surface moisture considering curing time and mix proportion

a.Figure 6 .
Figure 6.Regression analyses on the influence of density and moisture on the thermal conductivity

Figure 7 .6
Figure 7. Best validation performance of the ANN Figure 8. Error histogram

Figure 9 .
Figure 9. Overall performance of the ANN