Professional Documents
Culture Documents
Prem2022 Gaussian Models For Bond Strength Evaluation of Ribbed Steel Bars in Concrete
Prem2022 Gaussian Models For Bond Strength Evaluation of Ribbed Steel Bars in Concrete
Gaussian models for bond strength evaluation of ribbed steel bars in concrete
DOI
10.12989/sem.2022.84.5.651
Publication date
2022
Document Version
Final published version
Published in
Structural Engineering and Mechanics
Citation (APA)
Prem, P. R., & Savija, B. (2022). Gaussian models for bond strength evaluation of ribbed steel bars in
concrete. Structural Engineering and Mechanics, 84(5), 651-664.
https://doi.org/10.12989/sem.2022.84.5.651
Important note
To cite this publication, please use the final published version (if applicable).
Please check the document version above.
Copyright
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent
of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Takedown policy
Please contact us and provide details if you believe this document breaches copyrights.
We will remove access to the work immediately and investigate your claim.
https://www.openaccess.nl/en/you-share-we-take-care
(Received July 25, 2022, Revised October 7, 2022, Accepted October 28, 2022)
Abstract. A precise prediction of the ultimate bond strength between rebar and surrounding concrete plays a major role in
structural design, as it effects the load-carrying capacity and serviceability of a member significantly. In the present study,
Gaussian models are employed for modelling bond strength of ribbed steel bars embedded in concrete. Gaussian models offer a
non-parametric method based on Bayesian framework which is powerful, versatile, robust and accurate. Five different Gaussian
models are explored in this paper-Gaussian Process (GP), Variational Heteroscedastic Gaussian Process (VHGP), Warped
Gaussian Process (WGP), Sparse Spectrum Gaussian Process (SSGP), and Twin Gaussian Process (TGP). The effectiveness of
the models is also evaluated in comparison to the numerous design formulae provided by the codes. The predictions from the
Gaussian models are found to be closer to the experiments than those predicted using the design equations provided in various
codes. The sensitivity of the models to various parameters, input feature space and sampling is also presented. It is found that
GP, VHGP and SSGP are effective in prediction of the bond strength. For large data set, GP, VHGP, WGP and TGP can be
computationally expensive. In such cases, SSGP can be utilized.
Keywords: bond strength; concrete; gaussian; modelling; steel bars
Table 1 Bond strength models complicated processes and many non-Gaussian processes
Reference Model may be modelled. In the context of supervised learning, it
Orangun et al. 3𝑐min 50𝑑𝑏 𝐴𝑠𝑡1 𝑓𝑦𝑡
offers a precise and reliable Bayesian framework. It yields
√𝑓𝑐 [1.2 + + + 500𝑠𝑑 ] linear and theoretically predictable answers. By assigning
(1977) 𝑑𝑏 𝑙𝑏 𝑏
Xu 0.7𝑐min 𝐴𝑠𝑡1
priors to the functions and choosing appropriate kernels, it
𝑓𝑡 [1.6 + + 20 𝑠𝑐 ] has benefit of introducing previously known information
(1990) 𝑑𝑏 min
1 into the system. It has ability to incorporate the uncertainty
(𝑓𝑐 )4
Zuo and 𝜋𝑑𝑏 𝑙𝑏
[59.8𝑙𝑏 (𝐶min + 0.5𝑑𝑏 ) + 2,350𝐴𝑏 ] + estimation in the deep training model, unlike other machine
Darwin 𝐶 (𝑓 )4
1
𝑁𝐴𝑠𝑡 1
learning approaches like linear regression, neural networks,
(2000) 0.1 𝐶max + 0.9 + 𝜋𝑑𝑐 𝑙 [31.14𝑡𝑟 𝑡𝑑 + support vector machines, etc. The output is the distribution
min 𝑏 𝑏 𝑛
3.99)](√𝑓𝑐 ) of the predicted variable and not just one single value.
2⁄ Other machine learning algorithms don’t include this data in
Harajli 𝐴 3
√𝑓𝑐 [0.78 [(𝑐min + 7 𝑆 𝑠𝑡𝑛) /𝑑𝑏 ] ] their learning processes. Additionally, the training data set is
(2004) 𝑖
1
used directly to get the hyper parameters of the Gaussian
1+
𝜏𝑐 1.85+0.024√ 𝑀 (0.88 + 0.12 𝐶𝑚𝑒𝑑 ) (1 +
𝑀 𝐶 models.
𝐴𝑠𝑡1 𝐴𝑏
min In the present study, the effect of bar diameter,
Esfahani and 0.015 ) proportion of splice length to bar diameter (ls/db), relative
𝐶min 𝑠
Kianoush 𝑓 rib area of reinforcement (Rr), compressive strength of
(2005) 𝑀 = cosh (0.0022𝑙𝑏 √3 𝑑𝑐 )
𝑏 concrete (fc), proportion of minimum cover to bar diameter
𝐶min /𝑑𝑏 + 0.5 (Cmin/db), transverse reinforcement index (Atrfy/sndb) and
𝜏𝑐 = 2.7 √𝑓
𝐶min /𝑑𝑏 + 3.6 𝑐 yield strength of rebar (fy) on bond strength is modelled
0.265√𝑓𝑐 (
𝑐min
+ 0.5) using Gaussian processes. The models are validated using
AS3600 (2018) 𝑑
experimental observations. The performance of the models
𝑐min = min(𝑐𝑠𝑜, 𝑐𝑏 )
(𝑓𝑐 )1/4
is also compared with various codal formulas. The
[59.9𝑙𝑏 (𝐶min + 0.5𝑑𝑏 ) + 2,400𝐴𝑏 ] + sensitivity analysis of the models is also carried out to study
𝜋𝑑𝑏 𝑙𝑏
ACI408R-03 𝐶max (𝑓𝑐 )1/4 𝑁𝐴𝑠𝑡 1 their robustness. The methodology used for the current
(2012) 0.1 𝐶 + 0.9 + [30.88𝑡𝑟 𝑡𝑑 +
min 𝜋𝑑𝑏 𝑙𝑏 𝑛 study is presented in Fig. 1. The comparison of the different
3)](√𝑓𝑐 ) Gaussian models is (Verma et al. 2023) given in Table 2.
𝑓 0.25 25 0.25 𝑙 0.25
CEB-FIB 54 (25𝑐 ) (𝑑 ) (𝑑𝑏 )[𝛼2 + 𝑘𝑚 𝐾𝑡𝑟 ]
𝑏 𝑏
(2013) 𝐶min 0.25 𝐶max 0.1
2. Data set preparation
𝛼2 = ( ) (𝐶 )
𝑑𝑏 min
CSA-A23.3 0.2768
[
𝐴𝑡𝑟 𝑓𝑦𝑡
+
𝑑𝑐𝑠
] √𝑓𝑐 The basis for variable selection, partitioning of the data
(2004) ∏4𝑖=1 𝑘𝑖 10.5𝑠𝑛𝑑𝑏 𝑑𝑏
and preliminary data analysis are presented in this section.
EC2 2.25ῃ1 ῃ2 𝑓𝑐𝑡𝑑
(2004) ∏5𝑖=1 𝛼𝑖 2.1 Selection of input feature space
𝐶min 15 𝐴𝑡𝑟
0.31[(0.318 + 0.795 ( + )) +
JSCE 𝑑𝑏 𝑠 𝑑𝑏 Based on the critical review of the literature and the
(2007) 𝑑𝑏 study of the models presented in Table1, the major
13.3 ] √𝑓𝑐 parameters affecting the bond strength are found to be db,
𝑙𝑠
Rr, fc, Cmin/db, Atrfy/sndb, ls/db and fy. The experimental
Wu and 2.5
observations are collected from the literature which have
√𝑓𝑐 ( )
Zhao (2013)
1+3.1 𝑒
−0.47((𝐶min/𝑑 )+33𝐾𝑡𝑟 )
𝑏 considered all the aforementioned parameters (Darwin et al.
1995, Esfahani and Rangan 1998, Zuo and Darwin 2000
Mousavi et al. 3.9 Harajli 2004, CEB-FIB 2013, Golafshani et al. 2014). The
√𝑓𝑐 ( )
(2017) −0.54((𝐶min/𝑑𝑏 )+(
𝑑𝑏
⁄𝑙 )) test specimen consisted of a concrete cube embedded with a
1 + 12 𝑒 𝑏
10 mm diameter steel rod embedded in the centre of the
Rockson et al.
-2.84 + 0.84 (
𝐶min
) + 1.44 √𝑓𝑐 cross section. The bond strength was then evaluated by
(2020) 𝑑𝑏 performing pull-out tests using standard testing procedure.
set are evaluated. After partitioning of the data, same It can be seen that the statistical properties of the data in all
statistical properties are evaluated for the training and three groups are close to each other.
testing data sets. The data is partitioned in such a way the
statistical properties of the training and testing data sets are 2.3 Data analysis
close to that of the complete data set. This ensures that the
models are trained for the complete range of the input A feature correlation heat map is plotted between the
parameters. This also makes sure that the prediction of the input and output feature space as shown in Fig. 2. A pair
bond strength is not requested for the parameters whose plot is shown in Fig. 3 to visualize the relation between
range exceed the value for which it was trained. The different variables. The inference from the same is as
statistical properties of the input and output parameters for follows -
training, testing and complete data set are given in Table 3. • It is observed that there is a strong negative correlation
654 Prabhat R. Prem and Branko Savija
between the ratio of splice length to bar diameter and the confinement effect.
the bond strength. This is true because the increase in • A positive correlation is observed between the
the ratio of splice length to bar diameter results in reinforcement index and bond strength. This is apparent
decrease of bond strength as per (ACI408R-03 2012) since well-confined steel bars with the stirrups are
and (JSCE 2007). expected to exhibit higher bond strength compared to
• A negative correlation exists between the bar diameter unconfined bars.
and the bond strength indicating that a smaller diameter • The bond strength has a positive correlation with the
bar has more bond strength compared to the bigger concrete compressive strength. This is in line with the
diameter bars. equations provided in various design codes (ACI408R-
• A positive correlation is observed between minimum 03 2012, CEB-FIB 2013, CSA-A23.3 2004, EC2 2004,
concrete cover to bar diameter ratio and bond strength. JSCE 2007).
As the ratio increases, the bond strength increases due to
Gaussian models for bond strength evaluation of ribbed steel bars in concrete 655
Table 3 Statistical properties of data set mean (M), standard Table 4 Error indices for the Gaussian models
error (SE), median (ME) ERROR INDICES
Complete data set S. NO MODEL
ME RMSE MAE R
M 28.05 0.10 1.56 30.24 19.12 51.24 638.40 5.34
1. GP 0.011 0.517 0.405 0.973
SE 0.58 0.00 0.04 1.08 2.67 1.69 9.96 0.12
2. VHGP 0.014 0.511 0.399 0.974
ME 25.40 0.09 1.50 24.00 9.18 40.14 555.66 5.24
MO 25.40 0.08 1.50 16.00 0.00 35.86 827.59 5.32 3. SSGP 0.011 0.508 0.400 0.974
SD 8.69 0.03 0.55 16.23 40.10 25.28 149.37 1.73 4. WGP 0.266 1.611 1.239 0.697
V 75.48 0.00 0.30 263.28 1608.14 639.20 22311.68 2.99 5. TGP 0.085 0.821 0.648 0.925
K 6.55 0.31 2.03 1.57 23.99 -0.03 -1.67 0.33
S 1.94 1.05 1.04 1.39 4.71 1.07 0.23 0.44
R 50.80 0.12 2.90 87.00 271.21 86.27 413.80 9.84 model is near to 45◦, indicating that the predicted values are
Min 12.70 0.06 0.60 7.00 0.00 24.10 413.79 1.52 quite close to those found in the trials. According to the
Max 63.50 0.18 3.50 94.00 271.21 110.37 827.59 11.36 figure, GP, VHGP, and SSGP perform better than WGP and
Training TGP in terms of bond strength prediction since they are
M 27.89 0.10 1.55 30.28 18.21 51.78 632.43 5.28 closer to the line of best fit. For WGP, it is discovered that
SE 0.64 0.00 0.04 1.22 2.67 1.91 11.07 0.12 there is the least connection between experimental and
ME 25.40 0.09 1.50 24.00 9.18 40.24 555.66 5.22 projected bond strength. The mean error (ME), mean
MO 25.40 0.08 1.50 16.00 0.00 35.86 827.59 5.32 absolute error (MAE), root mean squared error (RMSE),
SD 8.73 0.03 0.52 16.61 36.28 25.87 150.10 1.63 and correlation coefficient (R), which are determined by the
V 76.26 0.00 0.27 275.85 1315.98 669.06 22530.47 2.66 following formulae, are used to compare the anticipated and
K 6.94 0.31 1.78 1.75 26.06 -0.16 -1.63 0.40 experimental results.
S 2.08 1.04 0.86 1.45 4.77 1.03 0.28 0.34 1
𝑀𝐸 = ∑𝑛𝑖=1 𝑦𝑖 − 𝑥𝑖 (1)
R 50.80 0.12 2.90 87.00 271.21 86.27 413.80 9.84 𝑛
Min 12.70 0.06 0.60 7.00 0.00 24.10 413.79 1.52 1
𝑀𝐴𝐸 = ∑𝑛𝑖=1|𝑦𝑖 − 𝑥𝑖 | (2)
Max 63.50 0.18 3.50 94.00 271.21 110.37 827.59 11.36 𝑛
Testing 1 2
M 28.78 0.10 1.63 30.08 23.18 48.82 665.17 5.60
𝑅𝑀𝑆𝐸 = √ ∑𝑛𝑖=1(𝑦𝑖 2 − 𝑥𝑖 ) (3)
𝑛
SE 1.34 0.01 0.10 2.28 8.50 3.53 22.62 0.33 𝑛(∑ 𝑥𝑖 𝑦𝑖 )−∑ 𝑥𝑖 ∑ 𝑦𝑖
ME 25.40 0.08 1.45 24.00 8.58 37.86 558.62 5.26 𝑅= (4)
√[𝑛 ∑ 𝑥𝑖 2 −∑ 𝑥𝑖 2 ][𝑛 ∑ 𝑦𝑖 2 −∑ 𝑦𝑖 2 ]
MO 25.40 0.08 2.13 24.00 0.00 36.21 827.59 4.55
SD 8.55 0.03 0.67 14.57 54.44 22.61 144.81 2.12 where yi are the predictions, xi are the experimental bond
V 73.16 0.00 0.44 212.31 2964.23 511.20 20969.75 4.50 strength and n is the number of samples. The error indices
K 5.86 0.34 1.82 -0.09 16.74 0.80 -1.90 -0.30 are shown in Table 4. The RMSE of the models follow the
S 1.39 1.12 1.31 0.99 4.14 1.29 0.04 0.52 order — WGP>TGP>GP>VHGP>SSGP. The RMSE and R
R 50.80 0.11 2.90 52.54 271.21 79.96 375.59 9.00 values for the models GP, VHGP and SSGP are close to
Min 12.70 0.07 0.60 12.00 0.00 27.97 452.00 1.92 each other. The WGP model is found to be least efficient in
Max 63.50 0.18 3.50 64.54 271.21 107.93 827.59 10.92 the prediction of bond strength compared to other Gaussian
models. This is presumed to be due to the limited number of
basic functions (five in this case) used for warping. Other
3. Results and discussions factors can be overfitting and non-modelling of additional
output noise after warping. In comparison to other models,
The bond strength is predicted in this section using GP, SSGP and VHGP are proven to have greater predicting
VHGP, SSGP, WGP, and TGP models. The models are first abilities when considering the total error indices.
trained using the training data set. The models thus
produced are then applied to the testing data set to predict 3.3 Residuals
the bond strength. Based on (i) error indices, (ii) residuals,
(iii) convergence, and (iv) computing time, the effectiveness The term “residual” refers to the discrepancy between
and capacity of the models to forecast bond strength are the actual and expected bond strength. The variability of the
assessed. The same’s specifics are expanded below. residuals aids in examining the model’s consistency in
correctly forecasting bond strength. Figure 6b displays the
3.1 Error indices distribution of the residuals for each Gaussian model as a
box plot. The 25th and 75thpercentiles are represented by
Fig. 4 illustrates the bond strength produced from the box’s outer corners, while the median is represented by
several Gaussian models with the experimental data. It is the red line in the centre. The outliers are identified by the
clear that the predictions of GP, VHGP, and SSGP are rather red symbols. For SSGP and WGP the changes in the
similar to the results of the trials. In Fig. 5, the black dots residuals are found to be least and most, respectively. Also,
indicate the sample of testing data, and the yellow line is the for SSGP and WGP, the mean residual is minimal and
line of best fit. The slope of the line of best fit for a decent maximum, respectively.
656 Prabhat R. Prem and Branko Savija
3.4 Computational time complete all of the tasks necessary for training the models.
The amount of CPU time used by various models is
The models’ computational time is stated in terms of displayed in Fig. 6(c). As can be seen, VHGP requires the
CPU time, which represents how long it takes the CPU to most computing time, whereas TGP requires the least. In
Gaussian models for bond strength evaluation of ribbed steel bars in concrete 657
(a)
Fig. 5 Continued
Fig. 7 Comparison of bond strength predicted from Gaussian models with those obtained using codal formulas for
testing data set
5.2 Sensitivity to size of input feature space case to study the sensitivity of the model to the number of
inputs. The inputs parameters that are omitted as selected
The results of the Gaussian models depend on the based on their correlation with the output. Therefore, in the
number of input parameters used for the modelling. The first case, Rr, which has the least correlation with bond
output of the Gaussian models might be less sensitive to strength, is removed. In the second case, the parameters
some of the input parameters that are used for the Rr,fy are removed. For third case, the parameters Rr,db,fy are
modelling. The omission of these inputs parameters from omitted while for the fourth case parameters Rr,db,fy,fc. The
the modelling may not significantly affect the output of the error indices are summarized in Table 6. In general, it is
Gaussian models. Four different cases are considered. The found that the RMSE and R values reduces with the
number of input parameters are reduced from 6 to 3. The decrease in the number of input parameters. This is
Gaussian models are then trained with the reduced number expected as some of the input parameters that are known to
of the inputs. The error indices are evaluated for each of the affect the bond strength have not been used for modelling.
Gaussian models for bond strength evaluation of ribbed steel bars in concrete 659
Table 7 Variation of error indices with the number of data Table 8 Error indices for the validation data set
set in training and testing ERROR INDICES
S. NO MODEL
Number of
Method ME RMSE MAE R ME RMSE MAE R
data points 1. GP 0.011 0.523 0.415 0.973
GP 0.063 0.574 0.464 0.955 2. VHGP 0.015 0.517 0.408 0.974
VHGP 0.075 0.572 0.461 0.955 3. SSGP 0.010 0.514 0.409 0.974
Testing 155,
SSGP 0.064 0.567 0.461 0.956 4. WGP 0.223 1.600 1.219 0.712
Testing 70
WGP 0.217 1.184 0.861 0.791 5. TGP 0.053 0.802 0.630 0.929
TGP 0.265 1.06 0.837 0.864
GP 0.137 0.69 0.551 0.931
VHGP 0.172 0.698 0.556 0.931 5.4 Sensitivity to external data set
Testing 145,
SSGP 0.084 0.67 0.522 0.934
Testing 70 The accuracy of the model is also studied with respect to
WGP 0.203 0.863 0.639 0.892
the external data set. A total of 40 samples which have
TGP 1.174 2.008 1.421 0.701 considered db, Rr, fc, Cmin/db, Atrfy/sndb, ls/db and fy as
GP 0.534 1.146 0.859 0.836 experimental parameters are collected from the literature
VHGP 0.531 1.139 0.857 0.838 and categorized as the validation data set (ACI408R-03
Testing 135, 2012, Golafshani et al. 2014, Cairns 2015). The trained
SSGP 0.287 1.09 0.815 0.823
Testing 70
WGP 2.071 7.073 2.251 -0.192 Gaussian models obtained in Section 3 are used for the
TGP 1.393 2.472 1.775 0.695 prediction of bond strength for the validation data set. The
error indices obtained using different Gaussian models for
GP 0.5 1.479 1.116 0.689
the validation data set are summarized in Table 8. The
VHGP 0.494 1.472 1.113 0.692
Testing 125, comparison between the experiment model results for the
SSGP 0.361 1.555 1.145 0.638 validation data set is given in Fig. 11. Based on the error
Testing 70
WGP 0.246 1.75 1.42 0.378 indices and correlation plots, GP, VHGP and SSGP have
TGP 9.099 11.162 0.4 0.423 comparable performance and are able to provide a good
prediction of the bond strength for validation data set. The
Gaussian models can sometime over-fit the data, especially
GP is found to be least sensitive to the number of input when the number of training data set is small. The over-
parameters. The error indices for WGP are found to not fitting means that the Gaussian models are trying to fit even
follow any trend. Therefore, one needs to be careful while the noise present in the output feature space. However, the
using WGP as the choice of input parameters can drastically case of over-fitting is not observed in the present study. The
effect the results of the model. good performance of the Gaussian models for the external
validation data set implies that there was no over-fitting of
5.3 Sensitivity to sampling the Gaussian models (Note that the error indices for the
validation data set are close to that for the testing data set).
The accuracy of the models also varies with the number For the case where, over-fitting of the models is observed, it
of data points used for the training of the model. Four can be avoided by choosing hyper parameters based on
different cases are considered herein, that is, the total data some prior knowledge. In case no prior knowledge is
set is split for training and testing as 155-70, 145-80, 135- available, the variance of the noise kernel can be increased
90 and 125-100. The error indices evaluated for each of the to avoid over-fitting.
cases are summarized in Table 7. It is observed that as we
decrease the number of training data points, the accuracy of
the models also decreases. Conversely, using more data 6. Conclusions
points for the training of the model results in better
prediction of the bond strength. Among all the Gaussian The development of an adequate model for the
models, GP, VHGP, SSGP are found to be perform prediction of bond strength is one of the active and
comparatively better than WGP and TGP with the less important research problems in structural engineering. In
training data. this paper, Gaussian models are adopted for the bond
In order to have a correlation coefficient more than 0.9, strength evaluation of the ribbed bars embedded in
it is advised to use at least 65% of the total data for the concrete. Gaussian models provide a Bayesian framework
training of the models. The RMSE values are found to with high predictive power. The input feature space
increase by a factor of almost 3 for GP, VHGP and SSGP consisted of bar diameter, relative rib area of the
when the number of training data points are reduced from reinforcement, ratio of minimum cover to bar diameter,
155 to 125. The performance of TGP degrades drastically ratio of splice length to bar diameter, transverse
with the reduction in the training data. The RMSE values reinforcement index, concrete compressive strength and
are found to increase from 1.06 to 11.162 while the R yield strength of the steel bar. The output was bond strength
values are reduced from 0.864 to 0.423 when the training obtained from the rod pull-out tests.
data is reduced from approximately 70% to 55% of the total Following conclusions are drawn from the present study
data set. • Five different Gaussian models were studied-Standard
660 Prabhat R. Prem and Branko Savija
Fig. 8 Distribution of the bond strength obtained from Gaussian models and various codes
results closer to the experiments. GP, VHGP, SSGP and training sample size. It was discovered that the models
TGP are found to provide better prediction than those GP, VHGP, and SSGP are capable of capturing the bond
obtained using codal formulas. strength’s sensitivity to different input factors The
• The sensitivity analysis of the models was carried out models performed well even with fewer input variables.
in terms of parameters, number of input features and The least sensitive model to the set of input variables
662 Prabhat R. Prem and Branko Savija
(e) TGP
Fig. 11 Comparison between experiments and model results for the validation data
Nomenclature confinement
𝜂𝑖 : concrete quality and rebar coefficients
𝑎 : hyperparameter for step size 𝜀 : constant power Gaussian noise
𝑎𝑟 , 𝑏𝑟 : amplitude parameters 𝜎2 : noise power
𝐴𝑏 : area considering single ribbed longitudinal bar 𝑣 : allowable covariance (signal power)
𝐴𝑠𝑡 : area considering all legs of the stirrups 𝛿 : Kroenke delta function
𝐴𝑠𝑡1 : area of stirrups considering one leg 𝜎𝑓 : length-scale for the input feature space
: area of individual stirrups passing potential 𝜎𝑛 : noise standard deviation
𝐴𝑡𝑟
zone of splitting failure in steel bar
𝐴𝑡𝑟 𝑓𝑦 ⁄𝑠𝑛𝑑𝑏 : transverse reinforcement index
𝑏 : hyper parameter for steepness
𝑐 : hyper parameter for position
𝑐𝑏 : bottom concrete cover
𝑐𝑠𝑖 :1/2 of the clear spacing between bars
𝑐𝑠𝑜 : side concrete cover
𝑐𝑠 : min (𝑐𝑠𝑜 ; 𝑐𝑠𝑖 + 6:35 mm)
𝑐min : min (𝑐𝑏 ; 𝑐𝑠 )
𝑐max : max (𝑐𝑏 ; 𝑐𝑠 )
𝑐𝑚𝑒𝑑 : average concrete cover
𝑐𝑠𝑘 : 2/3 of the gap between ribbed bars
𝑑𝑏 : diameter of ribbed bar
𝑑𝑐𝑠 : min (𝑐𝑠𝑜 + 0:5𝑑𝑏 ; 𝑐𝑏 + 0:5𝑑𝑏 ; 𝑐𝑠𝑘 )
𝐸 : Expectation
𝑓(𝑥) : latent function of the inputs
𝑓𝑐 : concrete compressive strength
𝑓𝑦 : yield strength of the spliced reinforcement
𝑓𝑦𝑡 : yield strength of the transverse reinforcement
: monotonic mapping function for WGP
𝑔(. , 𝜓)
parameterized by scalar inputs
𝑘(. ) : covariance or kernel function
: coefficients for bar location, coating, concrete
𝑘𝑖
density and rebar size
𝑘𝑚 : connement efficiency
𝑘𝑡𝑟 : stirrup index of connement
𝐾𝑓𝑓 : covariance matrix
𝐾𝐿 : Kullback-Leibler divergence
: number of basis function used for the
𝐿
parameterization of function g
𝑙𝑠 : splice length
𝑙𝑏 : bond length
𝑚 : count of basic functions
𝑛 : count of tension bars confined with stirrups
: count of transverse stirrups confined within
𝑁
the bond length
𝑝𝑠(𝑠) : probability measure
𝑞(. ) : variational density
𝑅𝑟 : relative rib area of the steel
𝑠𝑟 : spectral frequency vector
𝑆(𝑠) : power spectrum in frequency domain
𝑠𝑟 : spectral frequencies
𝑠 : gap between the transverse reinforcement
𝑠𝑖 : slip
𝑢𝑏 : bond strength
𝑥 : observed input
𝑥∗ : forecast input
𝑦 : observed output
𝑦∗ : forecast output
𝑧𝑖 : latent space for WGP
𝛼𝑖 : coefficients for bar shape, concrete cover and