Professional Documents
Culture Documents
a r t i c l e i n f o a b s t r a c t
Article history: In self-compacting concrete (SCC), admixtures and other modifiers are commonly used to decrease
Available online 2 March 2022 bleeding and segregation. Higher air content may be necessary to improve the mixture’s ability to with-
stand freezing and thawing. Based on the suggestions of the AFGC and EFNARC, several SCC mixes were
Keywords: created in this study. There was an air-entrainment admixture (AEA) applied in various percentages.
Self-compacting concrete Experimental studies have been performed to examine the mechanical characteristics of SCC’s, such as
Compressive strength sonic velocity (at 1 and 7 days) and compressive strength (at 1, 7, and 28 days) using different air content.
Multiple linear regression
This mechanical characterization is used taking into account void ratio and water absorption.
Support vector regression
Decision tree regression
In this sense, the objective of this work is to predict the compressive strength at 28 days from the
Random forest regression intrinsic parameters such as void ratio, water absorption, and the mechanical responses, at a young
age, such as sonic velocity and compressive strength (at 1 and 7 days).
To do this, we used the most famous machine learning algorithms to know: Multiple Linear Regression
(MLR), Random Forest Regression (RFR), Decision Tree Regression (DTR) and Support Vector Regression
(SVR). Finally, the best-proposed model is given on the basis of statistical comparison between the differ-
ent used algorithms.
Copyright Ó 2022 Elsevier Ltd. All rights reserved.
Selection and peer-review under responsibility of the scientific committee of the Third International Con-
ference on Aspects of Materials Science and Engineering.
https://doi.org/10.1016/j.matpr.2022.02.487
2214-7853/Copyright Ó 2022 Elsevier Ltd. All rights reserved.
Selection and peer-review under responsibility of the scientific committee of the Third International Conference on Aspects of Materials Science and Engineering.
Yousef El Asri, Mouhcine Ben Aicha, M. Zaher et al. Materials Today: Proceedings 57 (2022) 859–866
statistical models developed to efficiently predict the long-term of samples. During modeling, each tree randomly selects fea-
mechanical resistance according to the parameters mentioned tures to divide the attributes of the internal nodes, thus forming
above. a forest. The result of the final output is the result of the com-
Consequently, the originality of this work lies in the use of the plete decision tree. For the regression algorithm, the final result
four most popular ML methods, namely: support vector regression of the prediction is the average value of the output results of all
(SVR), decision trees regression (DTR), random forest regression trees [24,25].
(RFR), and multiple linear regression (MLR). Decision Tree Regression (DTR) model: In a ML system, a DT
The ultimate purpose of the proposed research is to find the algorithm separates the data into sub-groups in a machine
best mathematical model, with the most representative statistical learning (ML) system. When it comes to classification and
coefficients, of the four Machine Learning methodologies stated regression problems, Decision Trees can be utilized to execute
previously in order to predict the compressive strength of an SCC. split tests in their core nodes and forecast the target example
A dataset of 100 samples with similar physical and chemical in their leaf nodes [26].
characteristics will be used to find the best ML models for forecast- Support Vector Machine (SVM) model: This model is a super-
ing SCC’s compressive strength at 28 days [20]. vised learning technique used for regression and classification
analysis. The data is represented as a point map, and the solu-
2. Machine learning algorithms tion is the hyperplane with the greatest potential deviation that
divides this data into two classes. In this interval, each point is
The study of computer algorithms that improve automatically represented by a support vector [27]. For more precision, the
based on previous experiences and data is commonly known as support vector regression (SVR) was used in this study.
machine learning [21]. Supervised Learning, Unsupervised Learn-
ing, and Reinforcement Learning are the three types of machine 3. Research methodology
learning described in the literature [22]. Because we already have
input and output data, we’ll be using supervised machine learning 3.1. Experimental data
in this paper. The following is a summary of the mathematical and
numerical features of the four ML algorithms utilized in this study: The data set used in this study is composed of 100 self-
compacting concrete mixtures [10,28,29]. The values of these data
Multiple Linear Regression (MLR) model: It is based on the lin- are shown in Appendix A, with 6 input data such as water absorp-
ear regression, to find the correlation between two variables tion, void ratio, sonic velocity at 1 day, sonic velocity at 7 days,
(one input x and one output y), under this form: y = ax + b compressive strength at 1 day, and compressive strength at 7 days.
where ‘‘b‘‘ is the intercept and ‘‘a” is the slope of the line. Sim- The output data is compressive strength at 28 days.
ilarly, if there are more than one input and output variable, we
name it Multiple Linear Regression [23]. 3.2. Data pre-processing
Random Forest Regression (RFR) model: This prediction
methodology addresses prediction problems based on the aver- In this step, the Min-Max scaler normalization techniques have
age of a large number of distinct trees;RFR uses the Bootstrap been applied only for the input data. However, the normalization
resampling technic to build a decision tree model for each set of the output data is not necessary since we have only one output
860
Yousef El Asri, Mouhcine Ben Aicha, M. Zaher et al. Materials Today: Proceedings 57 (2022) 859–866
data for prediction (compressive strength at 28 days). This normal- All these steps are summarized in the following diagram
ization is very useful to facilitate the performance of modeling (Fig. 3).
since we have measures that have different units, the goal is to For evaluating and developing ML models, the results of this
homogenize all the input data between 0 and 1. study are obtained using Matplotlib, Numpy, Pandas, and Scikit-
The input data normalized is presented in Fig. 1. The calculation learn, which are libraries of open-source Python.
of this normalization is done using the Min-Max scaler of prepro-
cessing from the Scikit-learn library of Python.
4. Results and discussion
861
Yousef El Asri, Mouhcine Ben Aicha, M. Zaher et al. Materials Today: Proceedings 57 (2022) 859–866
Table 1
Data statistics: count, mean, std, min and max.
Input Output
Water Void Sonic velocity at Sonic velocity at Compressive strength at Compressive strength at Compressive strength at
absorption (%) ratio (%) 1 day (m/s) 7 days (m/s) 1 day (MPa) 7 days (MPa) 28 days(MPa)
Count 100 100 100 100 100 100 100
Mean 7.09 13.50 3762.82 4089.11 24.23 36.40 42.31
Std 3.64 6.63 373.62 420.02 10.77 13.07 12.40
Min 1.74 3.60 3286 3562 12.20 20.30 26.20
Max 12.72 24.35 4444 4790 44.00 57.00 65.60
Table 2
Results of test metrics of the variousML algorithmsto predictthe compressive strength at 28 days.
4.1. Prediction results performing model is RFR (with a large value of mean R2 and the
lowest errors), the second is DTR followed by MLR and lastly we
According to the ML algorithms discussed previously (MLR, RFR, find the SVR model.
DTR and SVR), the compressive strength of SCC (at 28 days) can be Therefore, the model which will be chosen for the prediction of
predicted using the results shown in Table 2. compressive strength at 28 days of SCC is Random Forest Regres-
The test metrics reported in Table 2 were obtained after divid- sion algorithm RFR (Fig. 5). This Figure analyzes the statistical coef-
ing our data into k fold cross-validation, and modeling each fold ficients mentioned previously. Thus, the comparison between the
using the 4 models for training and testing. true experimental and the predicted values of the adopted algo-
Table 2 shows that the mean R2 value and that of each fold and rithm model is presented in Fig. 6. The two graphs show that the
model exceed 0.99, which means that the four models predict com- predicted results and the experimental results are almost identical
pressive strength with high performance. with a higher correlation coefficient R2 = 0.9997 and minimal error
To choose the best-performing model, compared to the other with MAE = 0.134 MPa and MSE = 0.041, which means that the
three models, the histogram of Fig. 4 clearly shows that the best model is very efficient.
862
Yousef El Asri, Mouhcine Ben Aicha, M. Zaher et al. Materials Today: Proceedings 57 (2022) 859–866
863
Yousef El Asri, Mouhcine Ben Aicha, M. Zaher et al. Materials Today: Proceedings 57 (2022) 859–866
distinguishes it from other models in terms of its capacity to learn Software, Visualization, Investigation. Hafidi Alaoui Adil: Visual-
from new data. ization, Supervision, Validation.
El Asri Yousef: Conceptualization, Methodology, Software, The authors declare that they have no known competing finan-
Writing – original draft. Ben Aicha Mouhcine: Data curation, Writ- cial interests or personal relationships that could have appeared
ing – review & editing, Investigation, Validation. Zaher Mounir: to influence the work reported in this paper.
864
Yousef El Asri, Mouhcine Ben Aicha, M. Zaher et al. Materials Today: Proceedings 57 (2022) 859–866
865
Yousef El Asri, Mouhcine Ben Aicha, M. Zaher et al. Materials Today: Proceedings 57 (2022) 859–866
866