민하은
(Ha-Eun Min)
1iD
김희선
( Hee Sun Kim)
2†iD
-
이화여자대학교 건축도시시스템공학과 대학원생
(Graduate Student, Department of Architectural and Urban Systems Engineering, Ewha
Womans University, Seoul 03760, Rep. of Korea)
-
이화여자대학교 건축도시시스템공학과 교수
(Professor, Department of Architectural and Urban Systems Engineering, Ewha Womans
University, Seoul 03760, Rep. of Korea)
Copyright © Korea Concrete Institute(KCI)
키워드
콘크리트, 고온, 혼화재, 압축강도, 머신러닝 기법
Key words
concrete, high temperature, supplementary cementitious material (SCM), strength, machine learning
1. Introduction
Concrete has been predominantly used in construction due to its cost-effectiveness
and durability. Although concrete is also known to be fire-resistant due to its low
conductivity, its strength is sensitive to temperature changes (Ma et al. 2015). Additionally, it varies depending on the supplementary cementitious material (SCM)
content, such as fly ash and slag, which are increasingly being used as partial replacements
for cement to reduce environmental impact (Ramzi and Hajiloo 2023). Therefore, accurately predicting strength changes in concrete with varying SCM content
when exposed to high temperatures is crucial for assessing its fire resistance performance.
Recently, new approaches such as machine learning (ML) have been introduced, to elucidate
and predict concrete behaviors. Various studies have attempted to use ML models to
predict concrete compressive strengths, capture sensitivity, and overcome the limitations
of traditional statistical analysis. Farooq et al. (2020) adopted gene expression programming (GEP) and random forest (RF) algorithms to predict
the compressive strength of high-strength concrete. Ahmad et al. (2021a) predicted the compressive strength of fly ash-based concrete using decision trees
(DT), ensemble bagging approach, GEP, and K-Fold cross-validation methods. In particular,
several studies (Li and Song 2022; Rathakrishnan et al. 2022; Sapkota et al. 2024; Vo et al. 2024) have used ensemble models to better predict the compressive strength of concrete
by combining several weak learners. Vo et al. (2024) utilized adaptive boosting (AdaBoost), gradient boosting regression trees (GBRT),
extreme gradient boosting (XGBoost), and category gradient boosting (CatBoost), and
highlighted XGBoost’s superior performance. Rathakrishman et al. (2022) predicted
and compared the compressive strength of concrete replaced with a high volume ground
granulated blast furnace slag using light gradient boosting machine (LGBM), CatBoost,
gradient boosting regressor (GBR), AdaBoost, and XGBoost methods. Later, Sapkota et al. (2024) proved that CatBoost was more suitable for predicting normal concrete compressive
strength than RF.
Compared to ML studies on predicting compressive strength of concrete under a normal
environment, relatively few research have been reported on concrete exposed to high
temperatures. Ahmad et al. (2021b) compared the predictive accuracy of individual ML models with those from ensemble
ML models in estimating concrete compressive strength at high temperatures. They employed
DT and artificial neural networks (ANNs) as the individual models and bagging regressors
and GBR as the ensemble models. Their findings revealed that the ensemble models had
better predictive accuracy than individual models. Similarly, Ahmad et al. (2021c) examined the changes in the compressive strength of concrete heated to high temperatures
using three ML methods: DT (individual model), and AdaBoost and RF (ensemble models).
They conducted comparative and sensitivity analyses, which revealed that AdaBoost
possesses the most effective predictive performance among the three (Ahmad et al. 2021c).
Despite the importance and necessity of research evaluating the fire resistance performance
of concrete, there are lack of studies about ML to assess the fire resistance of structures.
The previous studies considered limited features, thus not enough to include a diverse
range of parameters influencing strength of fire damaged concrete. Moreover, comparative
studies for ensemble models have not been sufficiently conducted, which makes hard
to adopt the most recent ML techniques in fire safety engineering field.
Hence, this study endeavored to propose a ML model the most suitable to predict the
compressive strength of concrete heated to high temperatures among various ensemble
models. The model was developed using diverse dataset encompassing different parameters
such as normal or high strength, mix ratios of admixture, heating temperatures, and
cooling period after heating. In addition to investigating predictive accuracy and
conducting sensitivity analyses, this study validated the proposed model by further
applying a new dataset to the model and comparing the predicted and experimental values.
Ultimately, this research can facilitate the use of ML models to predict the fire
resistance of concrete structures by estimating changes in the compression strength
of concrete exposed to high temperatures.
2. Machine Learning Algorithms
Among many ML models, one notable type is the ensemble model, which outperforms individual
models by combining several weak learners (Feng et al. 2020; Ahmad et al. 2021a). The ensemble model increases the probability of creating a learner that shows better
generalization performance by combining multiple individual models. There are two
methods in an ensemble model: boosting and bagging. The boosting method sequentially
trains models and reduces errors by assigning weights to misclassified instances.
It involves adjusting the sample weights of the next classifier’s training data based
on the previous classifier’s learning results. Each classifier performs random sampling
with replacement. Representative boosting algorithms include gradient boosting regressor
(GBR), extreme gradient boosting regressor (XGBR), and categorical gradient boosting
(CatBoost). Bagging, short for bootstrap aggregating, is a parallel-type ensemble
method that creates each subset by randomly sampling the data with replacement (bootstrap)
and combines the predictions from these individual models (aggregating). Random forest
(RF) and extra trees (ET) are example models of bagging. In this study, five commonly
employed ensemble models are chosen: GBR, XGBR, CatBoost, RF, and ET, which are explained
in the following sections.
2.1 Boosting
2.1.1 Gradient boosting regressor (GBR)
GBR is based on a boosting algorithm and compensates for errors by assigning weights
to errors in the previous tree and employs a slope-down method for weight updates.
Although its sequential learning process can lead to slower processing, GBR can effectively
handle various optimized objective functions.
2.1.2 Extreme gradient boosting regressor (XGBR)
XGBR has been developed to overcome the boosting algorithm’s speed limitations and
demonstrate faster processing by supporting parallel processing. It contains a mechanism
known to regulate overfitting and shows outstanding predictive capabilities in both
classification and regression.
2.1.3 Categorical gradient boosting (CatBoost)
CatBoost can manage categorical variables effectively through ordered boosting and
efficient encoding techniques of target, mean, and response encoding. Recently, the
model has been widely used because of the relatively low computation time without
compromising accuracy and optimized hyperparameters.
2.2 Bagging
2.2.1 Random forest (RF)
RF comprises DT and is capable of resolving overfitting issues occurred in a single
decision tree by utilizing multiple trees. Each RF tree generates varied predictions
due to randomness, enabling the model to learn from diverse perspectives and improve
its predictions, which ultimately enhances its generalization performance. Additionally,
by applying randomization to ensemble method bagging, the forest becomes strong when
there are noisy data.
2.2.2 Extra trees (ET)
ET, a variation of the RF model, randomly selects features for splitting nodes, while
RF typically chooses the optimized split nodes from the given features. In other words,
ET incorporates greater randomness than RF at greater executed speed. The model also
reduces bias by utilizing the entire original dataset without relying on Bootstrap,
a key characteristic of RF.
3. Database Description
3.1 Data collection
To predict the fire resistance performance of concrete using an ML algorithm, data
were collected experimentally as well as from the available literature. For the experiments,
test variables were designed with different water-to-binder ratio and cement-to-admixture
ratio, as listed in Table 1.
Cylindrical specimens with a diameter of 100 mm and a height of 200 mm were manufactured
and cured underwater for 28 days at room temperatures. Then, the cured specimens were
preheated at a constant temperature of 100 °C inside a heating muffle furnace (Nabertherm,
Germany) to prevent spalling by evaporating the moisture inside the specimens. Thereafter,
the specimens were heated at 200 °C, 500 °C, or 800 °C for 3 hours to ensure that
the target temperature was completely transferred through the specimens for a sufficient
duration. The rate of temperature increase was controlled not to be faster than 4.4
°C/min during heating to prevent spalling. After heating, the specimens remained in
the heating chamber for 24 hours until the temperature is slowly cooled down. Then,
compressive strength and strains were measured using a loading machine and compressometer,
respectively. Compressive strength tests were conducted two or three times for each
test variable until consistent test results were obtained (Chun et al. 2023). Previous study (Chun et al. 2023) also reported strength changes of the concrete from 1 day to 90 days after the heating,
and the heated specimens showed minimal strength decrease until 30 days after the
heating. Therefore, the change of strength within 24 hours after the heating could
be neglected, considering amount of the strength changes within 30 days after the
heating and consistent strength test conditions that were conducted on the specimens
placed on the heating chamber for 24 hours after the heating. Figs. 1(a)~(c) are the photographs of manufacturing specimens as well as the heating and loading
tests. Fig. 2 indicates that as the temperature rises from 200 °C to 800 °C, the compressive strengths
of the concrete decrease significantly. In order to investigate the effect of mix
ratios on amount of strength reduction due to heat, residual strength ratios were
calculated as strengths of heated specimens divided by those of unheated specimens
and illustrated in Figs. 3(a)~(c).
Fig. 3(a) illustrated residual strength ratios of pure cement, cement-fly ash, and cement-fly
ash-slag composite concretes at 200 °C, 500 °C and 800 °C. It was interesting to note
that the concrete with same amount of fly ash and slag showed the better residual
strength ratio than the pure cement and the cement with fly ash concrete when exposed
to 500 °C and 800 °C. In contrast, Fig. 3(b) showed that within the fly ash and slag composite concretes, concretes having different
mix ratios of fly ash and slag portions showed the larger strength reduction even
than the pure cement and the cement with fly ash concrete. Therefore, the experimental
results showed that the strength reduction ratios varied depending on mix ratios.
However, the relations between fly ash and slag are not clear and the more data need
to be analyzed to generalize the effect of mix ratios of admixtures on strength reduction
due to heat. When comparing residual strength ratios of normal (NF3S1) and low cement
(LF3S1) concretes, the higher residual strength ratios were obtained from the low
cement concrete with fly ash and slag composite (LF3S1) at all the tested temperatures
as shown in Fig. 3(c).
Even though there is a significant effect of mix proportions such as water-to-binder
ratio or cement-to- admixture ratio on strength of heated concrete, it is very hard
to predict because of multiple and complex influencing parameters. Given the difficulty
in statistically analyzing the trends in strength changes according to the heat exposure
and mix proportions of concrete, it is helpful to utilize ML for predicting the compressive
strength of fire damaged concrete. Then, ML algorithm needs input data imported from
the experimental results. The data were categorized by cement (kg), fly ash (cement-
to-fly ash ratio, %), slag (cement-to-slag ratio, %), W/B (water-to-binder ratio,
%), temp (heating temperature, °C), strength (compressive strength, MPa), and time
(cooling period after heating, d), in order to consider the influencing factors and
the expected output (strength). Because of a limited number of the experimental results
obtained from our research group, additional data on concrete strength following exposure
to high temperatures for different mix proportions were collected from published literature
(Yeh 1998; Lee et al. 2002; Kim et al. 2012; Lee et al. 2012; Ahmad et al. 2021b; Song et al. 2021) and added to the ML dataset. Therefore, a total of 738 datasets, collected from experiments
and the literature, are used to train and test the ML algorithms.
Fig. 1 Photographs of manufacturing specimens, heating, and loading test
Fig. 2 Compressive strength of the heated specimens
Fig. 3 Comparison of residual strength ratio of the heated specimens
Table 1 Mix proportion of the tested specimens
|
Mix ID
|
Cement
(kg/m3)
|
Water
(kg/m3)
|
Water-to- binder ratio (%)
|
Cement-to-
admixture ratio (%)
|
Fly ash
(kg/m3)
|
Slag
(kg/m3)
|
Fine aggregate
(kg/m3)
|
Coarse aggregate
(kg/m3)
|
PP fiber
(vol%)
|
Super-
plasticizer
(kg/m3)
|
|
NF0S0
|
604
|
157
|
26
|
0
|
0
|
0
|
681
|
921
|
0.15
|
6.04
|
|
NF4S0
|
484
|
157
|
26
|
20
|
120
|
0
|
662
|
895
|
|
NF2S2
|
484
|
157
|
26
|
20
|
60
|
60
|
666
|
909
|
|
NF3S1
|
484
|
157
|
26
|
20
|
90
|
30
|
663
|
904
|
|
NF1S3
|
484
|
157
|
26
|
20
|
30
|
90
|
669
|
913
|
|
LF4S0
|
300
|
140
|
31.2
|
40
|
100
|
100
|
606
|
1,055
|
|
LF2S2
|
300
|
140
|
31.2
|
40
|
200
|
0
|
617
|
1,073
|
|
LF3S1
|
300
|
140
|
31.2
|
40
|
150
|
50
|
611
|
1,063
|
3.2 Data preprocessing
Machine learning modeling was performed using Scikit- Learn, a Python-based library
widely used for machine learning analysis. The dataset was split into training and
testing sets, with proportions of 30 % and 70 %, respectively. Additionally, through
the data scaling standardization process, the mean of each variable was set to 0,
and the variance was set to 1, reducing the impact of scale differences between variables.
4. Model selection in machine learning
4.1 Model evaluation indices
The coefficient of determination (R2), mean absolute error (MAE), and root mean square error (RMSE) are utilized as indicators
in this study to evaluate the prediction performance of the models and choose the
best model. R2 is a variance-based indicator ranging from 0 to 1, and a value closer to 1 indicates
a higher prediction accuracy. MAE is the average of the absolute value of the difference
between actual and predicted values, whereas RMSE is the square root of the mean squared
differences between actual and predicted values. For both MAE and RMSE indices, the
lower values indicate higher accuracy. The formulations for obtaining evaluation indices
are presented in Eqs. (1)~(3), where N is the number of samples, y is the actual value, $\overline{y}$ is the mean
of the actual values, and $\hat{y}$ is the predicted value (Li and Song 2022).
4.2 Model evaluation and selection
After the preprocessing of input data and default parameter tuning through PyCaret,
the evaluation indices of R2, RMSE, and MAE values were collected from various ML models. Among many other models,
top five ML models having the highest prediction accuracy were found to be GBR, XGBR,
CatBoost, RF, and ET as listed in Table 2.
All evaluation indicators: R2, RMSE, and MAE, revealed that the accuracy level of the performance appeared to be
the best in the CatBoost model. For example, the value of R2 in the CatBoost model was 0.8356, whereas those from the other four models were 0.8196
(RF), 0.8146 (ET), 0.8097 (XGBR), and 0.8086 (GBR). The result also agreed with the
findings from Sapkota et al. (2024), where CatBoost was found to be more suitable for predicting normal concrete compressive
strength than RF.
To improve prediction accuracy, hyperparameter tuning was conducted by selecting and
implementing optimized parameters to the five ML models. These parameters, therefore,
are different from those determined automatically through PyCaret. In all models except
RF, the R2, RMSE, and MAE values improved through the hyperparameter tuning, and the CatBoost
model demonstrated better accuracy than the other models regardless of the evaluation
index types. Table 3 indicates that the R2 and RMSE values for the test dataset were best in the order of CatBoost, XGBR, GBR,
ET, and RF, whereas the MAE value was found to be best in order of CatBoost (4.7733),
ET (4.9702), GBR (5.0092), XGBR (5.2126), and RF (5.4410). It was interesting to note
that, CatBoost had the best performance in all cases, and the difference between the
bagging and boosting methods was marginal. Among the bagging models, ET was better
than the RF model.
Additionally, a comparative visualization of the actual experimental values and predictions
obtained from the five models is illustrated in Figs. 4(a)~(e). As shown in Figs. 4(a) and (b), regarding the CatBoost and ET models, the marks denoting the experimental results
(in blue and red) are hardly visible because they are overlapped with the prediction
marks (in green and yellow). However, in other models (Figs. 4(c)~(e)), the marks denoting experimental data (in blue and red) and predictions (in green
and yellow) are clearly visible, as they are relatively far apart.
Fig. 4 Comparisons between the actual dataset and predicted results of the five models in
the training and testing phases
Table 2 Evaluation results of different ML models
|
Model
|
Type
|
R2
|
RMSE
|
MAE
|
|
CatBoost
|
Boosting
|
0.8356
|
7.7724
|
5.2003
|
|
RF
|
Bagging
|
0.8196
|
8.1560
|
5.3934
|
|
ET
|
Bagging
|
0.8146
|
8.2800
|
5.2355
|
|
XGBR
|
Boosting
|
0.8097
|
8.4026
|
5.4116
|
|
GBR
|
Boosting
|
0.8086
|
8.3997
|
5.9740
|
Table 3 Evaluation results of hyperparameter tuning
|
Model
|
R2
|
RMSE
|
MAE
|
|
Train
|
Test
|
Train
|
Test
|
Train
|
Test
|
|
CatBoost
|
0.9818
|
0.8411
|
2.5911
|
7.6400
|
0.6685
|
4.7733
|
|
XGBR
|
0.9644
|
0.8337
|
3.6267
|
7.8148
|
2.1447
|
5.2126
|
|
GBR
|
0.9755
|
0.8220
|
3.0065
|
8.0855
|
1.6812
|
5.0092
|
|
ET
|
0.9818
|
0.8182
|
2.5911
|
8.1714
|
0.6685
|
4.9702
|
|
RF
|
0.9623
|
0.8020
|
3.7330
|
8.5271
|
2.2212
|
5.4410
|
5. Model performance analysis
5.1 Feature importance
A feature importance analysis was conducted among the five ML models, and the results
are illustrated in Fig. 5. All models indicated similar patterns; specifically, the water- to-binder ratio
and heating temperature were identified as the most influential factors affecting
the strength of heated concrete. Meanwhile, the cooling period had the least influence
on strength. Between the admixtures, fly ash exhibited a slightly stronger influence
than slag. Interestingly, the CatBoost model identified heating temperature as slightly
more important than the water-to-binder ratio, but the difference was small; and the
ranking of the remaining features was similar to that in other models.
In previous sensitivity analyses (Ahmad et al. 2021b; Ahmad et al. 2021c), cement was the most influential factor, followed by fly ash. Meanwhile, the effect
of heating temperature appeared less substantial than that of other factors. Unlike
the aforementioned research, this study recognized the contribution of temperature
to the strength of the heated concrete. This may explain why the CatBoost model had
a higher prediction accuracy than the other models.
Fig. 5 Feature importance for the five ML models
5.2 Validation of CatBoost and ET models
In this section, a new dataset that had not been used in training or testing was applied
to the two ML models, CatBoost and ET models, which had shown the highest accuracy
overall and were the best among the bagging models, respectively. This was done not
only to validate the prediction accuracy of the models, but also to demonstrate the
applicability of these models to wide ranges of concrete mixtures. The new dataset
was adopted from the work of Khan et al. (2013). They manufactured cube-shaped concrete specimens using four different mix ratios
by varying the ratio of cement to fly ash and water to binder. The compressive strengths
of the concrete were measured after being exposed to temperatures ranging from 100
°C to 800 °C with 100 °C intervals.
Out of four test variables of concrete mixes, only three mixes listed in Table 4 were used as input data of the ML models. Mix 2 of Khan et al. (2013) was excluded due to its ratios of cement to binder or water to binder being identical
to those of Mix 1, Mix 3, and Mix 4. The experimental results of the compressive strengths
were varied as 13 MPa, 39 MPa, and 51 MPa in Mix 1, Mix 3, and Mix 4 at room temperature,
respectively.
The compressive strengths predicted from the ML models were converted to cubic strength
using Eq. (4) (Tam et al. 2017), because the experimental results were based on cubic strength.
Fig. 6 illustrates the results of concrete strength at temperatures of 20 °C, 200 °C, 500
°C, and 800 °C predicted using the CatBoost and ET models (in black), compared to
experimental results (in gray). It is found from Fig. 6(a) that the trend of strength changes with temperature and mix ratio is similarly observed
in predictions using the CatBoost model. As temperature changed, both experimental
results and prediction from CatBoost model showed that the strengths slightly increased
at 200 °C but decreased significantly at 500 °C and 800 °C. Furthermore, the strength
differences according to the mix ratios were also predicted similarly to the experimental
values when using the CatBoost model. At 20 °C and 200 °C, the compressive strength
was highest for Mix 4 (circle mark), which has a low water-to-binder ratio, followed
by Mix 3 (square mark) with a low admixture substitution rate, and then Mix 1 (triangle
mark). Notably, as the temperature increased, Mix 4 exhibited a greater reduction
in strength compared to Mix 3, resulting in Mix 4 having lower compressive strength
than Mix 3 at 800 °C. This reversal in strength order at 800 °C was also accurately
predicted by the CatBoost model.
However, the bagging method, ET, predicted only minimal changes in strength at temperatures
of 20~500 °C, while experimental results and the CatBoost demonstrated relatively
large strength changes due to heat exposure. Only Mix 4 demonstrated strength predictions
that closely matched the experimental results of the ET model, which may be due to
its water-to-binder ratio being similar to that of the trained data in the ML model.
Therefore, it was consistently revealed that CatBoost was the most suitable method
for predicting the strength of heated concrete. The validated conditions were across
normal and high- strength concrete using cement alone, with fly ash or slag as admixtures
and across heating temperatures of 200 °C, 500 °C, and 800 °C.
Furthermore, there is a limitation such that the strength predicted from the ML models
were compared to the ones obtained from cubic specimens, while the developed ML models
targeted to predict strength of cylinderical shaped specimens. Even though the strengths
predicted from the ML models were converted to the strength of cubic specimens by
Eq. (4), the converted strength were approximated which might results inaccurate prediction.
In addition, it was found that the ML model predicted strength changes of concrete
at high temperatures more reasonably, rather than the strength at room temperature.
This might be because trained data had wide ranges and types of parameters but amount
of the trained data was less than needed.
Fig. 6 Experimental and predicted compressive strength of concrete exposed to high
temperature
Table 4 New dataset of mix proportions, adopted from Khan et al. (2013)
|
Mix ID
|
Cement
(kg/m3)
|
Fly ash
(kg/m3)
|
Water-to-binder ratio
|
|
Mix 1
|
136
|
204
|
0.45
|
|
Mix 3
|
204
|
136
|
0.45
|
|
Mix 4
|
198
|
198
|
0.35
|
6. Conclusion
This study utilized five ensemble machine learning algorithms-ET, RF, GBR, XGBR, and
CatBoost, known as ensemble models-to predict the compressive strength of concrete
subjected to high temperatures, considering the effects of SCM content. The most suitable
ensemble model was proposed by assessing and comparing predictive accuracies. In addition,
new input data were applied to the two most accurately predictive ML models, CatBoost
and ET, to validate the models’ performance. As a result, the CatBoost model was found
to be the most suitable for predictions, and it is expected that, with further development
through an expanded database, it could be utilized as a method for evaluating the
fire resistance performance of concrete in the future. The following are the detailed
findings of this study.
ET, RF, GBR, XGBR, and CatBoost models exhibited reasonable prediction performance;
their R2 value ranged from 0.8 to 0.9, with CatBoost standing out for its exceptional performance.
CatBoost demonstrated higher accuracy than the other models across all indicators
including R2, RMSE, and MAE. Minimal differences were observed between bagging and boosting methods,
but regarding the bagging method, ET outperformed the RF model.
The evaluation of the key factor influencing compressive strength of fire damaged
concrete highlighted the differences between CatBoost and the other four models. In
CatBoost, heating temperature was identified as the most influential factor, followed
by the water-to-binder ratio. Conversely, in the other four models, the order was
reversed, indicating a higher priority for the water-to- binder ratio over the heating
temperature.
When predicting strength from a new input that had not been used for training or testing,
CatBoost demonstrated a better prediction accuracy than the ET model. The CatBoost
model’s prediction of strength changes with heating temperatures and mix ratios well
captured experimental results, which showed significant strength increase from 20
°C to 200 °C and decrease from 200 °C to 800 °C.
Funding
This research was supported by the Basic Science Research Program funded by the
National Research Foundation of Korea (NRF) (NRF-2021R1F1A1051300).
Acknowledgement
The authors are grateful for the help of Chae-eun Lee and Sun-young Cha, undergraduate
students in the department of architectural and urban systems engineering, Ewha Womans
University, in conducting experiments. We are also grateful for the help of Jae-sun
Kwon, an undergraduate student in the department of architectural and urban systems
engineering, Ewha Womans University, in initializing machine learning codes.
References
Ahmad, A., Farooq, F., Niewiadomski, P., Ostrowski, K., Akbar, A., Aslam, F., and
Alyousef, R. (2021a) Prediction of Compressive Strength of Fly Ash Based Concrete
Using Individual and Ensemble Algorithm. Materials 14(4), 794.
Ahmad, A., Ostrowski, K. A., Maślak, M., Farooq, F., Mehmood, I., and Nafees, A. (2021b)
Comparative Study of Supervised Machine Learning Algorithms for Predicting the Compressive
Strength of Concrete at High Temperature. Materials 14(15), 4222.
Ahmad, M., Hu, J. L., Ahmad, F., Tang, X. W., Amjad, M., Iqbal, M. J., Asim, M., and
Farooq, A. (2021c) Supervised Learning Methods for Modeling Concrete Compressive Strength
Prediction at High Temperature. Materials 14(8), 1983.
Chun, Y., Kwon, J., Kim, J., Son, H., Heo, S., Cho, S., and Kim, H. (2023) Experimental
Investigation of the Strength of Fire-Damaged Concrete Depending on Admixture Contents.
Construction and Building Materials 378, 131143.
Farooq, F., Nasir Amin, M., Khan, K., Rehan Sadiq, M., Javed, M. F., Aslam, F., and
Alyousef, R. (2020) A Comparative Study of Random Forest and Genetic Engineering Programming
for the Prediction of Compressive Strength of High Strength Concrete (HSC). Applied
Sciences 10(20), 7330.
Feng, D. C., Liu, Z. T., Wang, X. D., Chen, Y., Chang, J. Q., Wei, D. F., and Jiang,
Z. M. (2020) Machine Learning-Based Compressive Strength Prediction for Concrete:
An Adaptive Boosting Approach. Construction and Building Materials 230, 117000.
Khan, M. S., Prasad, J., and Abbas, H. (2013) Effect of High Temperature on High-Volume
Fly Ash Concrete. Arabian Journal for Science and Engineering 38, 1369-1378.
Kim, J. B., Shin, K. S., and Park, K. B. (2012) Mechanical Properties of Ultra High
Strength Concrete Using Ternary Blended Cement. Journal of the Korea Institute for
Structural Maintenance and Inspection 16(6), 56-62. (In Korean)
Lee, B. S., Jun, M. H., and Lee, D. H. (2012) The Effect of Mixing Ratio of Blast
Furnace Slag and Fly Ash on Material Properties of 80MPa High Strength Concrete with
Ternary Cement. LHI Journal of Land, Housing, and Urban Affairs 3(3), 287-297. (In
Korean)
Lee, D. H., Seo, D. H., Jun, P. H., Paik, M. S., Lim, N. G., and Jung, S. J. (2002)
The Experimental Study on High Strength Concrete of High Volume Fly-Ash. KCI 2002
fall Conference. Korea Concrete Institute (KCI), 14(2), 275-280. (In Korean)
Li, Q. F., and Song, Z. M. (2022) High-Performance Concrete Strength Prediction Based
on Ensemble Learning. Construction and Building Materials 324, 126694.
Ma, Q., Guo, R., Zhao, Z., Lin, Z., and He, K. (2015) Mechanical Properties of Concrete
at High Temperature-A review. Construction and Building Materials 93, 371-383.
Ramzi, S., and Hajiloo, H. (2023) The Effects of Supplementary Cementitious Materials
(SCMs) on the Residual Mechanical Properties of Concrete after Exposure to High Temperatures.
Buildings 13(1), 103.
Rathakrishnan, V., Bt. Beddu, S., and Ahmed, A. N. (2022) Predicting Compressive Strength
of High-Performance Concrete with High Volume Ground Granulated Blast-Furnace Slag
Replacement Using Boosting Machine Learning Algorithms. Scientific Reports 12(1),
9539.
Sapkota, S. C., Saha, P., Das, S., and Meesaraganda, L. P. (2024) Prediction of the
Compressive Strength of Normal Concrete Using Ensemble Machine Learning Approach.
Asian Journal of Civil Engineering 25(1), 583-596.
Song, H., Ahmad, A., Farooq, F., Ostrowski, K. A., Maślak, M., Czarnecki, S., and
Aslam, F. (2021) Predicting the Compressive Strength of Concrete with Fly Ash Admixture
Using Machine Learning Algorithms. Construction and Building Materials 308, 125021.
Tam, C. T., Babu, D. S., and Li, W. (2017) EN 206 Conformity Testing for Concrete
Strength in Compression. Procedia Engineering 171, 227-237.
Vo, T. C., Nguyen, T. Q., and Tran, V. L. (2024) Predicting and Optimizing the Concrete
Compressive Strength Using An Explainable Boosting Machine Learning Model. Asian Journal
of Civil Engineering 25(2), 1365-1383.
Yeh, I. C. (1998) Modeling of Strength of High-Performance Concrete Using Artificial
Neural Networks. Cement and Concrete Research 28(12), 1797-1808.