Table 1. Conventional prediction models for the estimation of Mr Values.

Table 2. Machine learning prediction models for the estimation of MR value

 Table 3 Loading sequence as per AASHTO T 307
Cyclic
Max. Cyclic
Confining Contact Stress
deviator deviator
stress stress Ratio
stress stress
No. of load (CSR) Total axial
Seq. No.
Appl. σ dmax stress (kPa)
σc =σ3 σdmax σcyclic 0.1σdmax
σc
kPa kPa kPa kPa --
0 750 103.4 103.4 93.1 10.3 206.8
1 100 20.7 20.7 18.6 2.1 0.90 41.4
2 100 20.7 41.4 37.3 4.1 1.80 62.1
3 100 20.7 62.1 55.9 6.2 2.70 82.8
4 100 34.5 34.5 31 3.5 0.90 69
5 100 34.5 68.9 62 6.9 1.80 103.4
6 100 34.5 103.4 93.1 10.3 2.70 137.9
7 100 68.9 68.9 62 6.9 0.90 137.8
8 100 68.9 137.9 124.1 13.8 1.80 206.8
9 100 68.9 206.8 186.1 20.7 2.70 275.7
10 100 103.4 68.9 62 6.9 0.60 172.3
11 100 103.4 103.4 93.1 10.3 0.90 206.8
12 100 103.4 206.8 186.1 20.7 1.80 310.2
13 100 137.9 103.4 93.1 10.3 0.68 241.3
14 100 137.9 137.9 124.1 13.8 0.90 275.8
15 100 137.9 275.8 248.2 27.6 1.80 413.7
 Table 4 Various descriptive statistics of the variables

Variable

Octahedral
Gravel

Stress
OM

Bulk
Stress
MDD D10 D30 D50 D60 CGP Cu Cc Fines Sand RAP sig1 sig3 sigd J2 MR
C

Statistic
(KN/
Paramete (%) (mm) (mm) (mm) (mm) (mm) -- -- (%) (%) (%) (%) (kPa) (kPa) (kPa) (kPa) (kPa) (kPa) (Mpa)
m3)
r
182.466
Mean 21.22 6.87 0.46 2.13 5.44 7.73 3.94 26.22 1.41 3.89 21.11 51.95 41.66667 7 73.08 98.45 41472.30 46.37 328.63 238.65
Standard 27.5478
Error 0.12 0.14 0.06 0.22 0.49 0.66 0.34 3.12 0.06 0.08 1.62 2.52 22.04793 9 11.57 16.85 10380.91 7.94 48.77 3.66

Median 21.16 6.75 0.28 1.35 5.19 7.03 3.85 20.61 1.40 4.00 19.75 53.63 50 172.3 68.90 93.10 33244.25 43.85 344.60 236.00

Mode 21.10 6.50 0.23 1.35 6.50 2.88 1.45 12.50 0.99 4.00 10.00 26.00 #N/A 206.8 20.70 62.00 #N/A 29.20 #N/A 149.00
Standard 106.692
Deviation 0.74 0.83 0.34 1.33 2.95 3.95 2.01 18.72 0.38 0.46 9.72 15.91 38.18813 5 44.81 65.28 40205.09 30.74 188.88 85.09
Sample 11383.2 #######
Variance 0.55 0.69 0.11 1.77 8.69 15.58 4.06 350.28 0.15 0.22 94.47 253.27 1458.333 9 2007.96 4260.93 # 945.25 35675.54 7239.86

Kurtosis 0.57 -0.14 0.46 -0.28 1.76 1.32 1.16 5.90 -0.83 1.43 -0.35 -0.90 #DIV/0! -0.20181 -1.46 0.50 0.19 0.50 -0.99 -0.74
0.58555
Skewness 0.45 0.47 1.37 0.96 0.99 0.86 0.84 2.12 0.43 1.38 0.73 -0.14 -0.93522 9 0.27 1.03 0.97 1.03 0.31 0.28
#######
Range 3.50 3.43 1.13 4.55 14.00 18.25 9.31 93.33 1.42 1.50 32.00 57.00 75 372.3 117.20 229.60 # 108.14 606.70 432.00

Minimum 19.80 5.20 0.15 0.45 1.00 1.75 0.84 6.68 0.77 3.50 10.00 25.00 0 41.4 20.70 18.60 2142.45 8.76 82.80 73.00
#######
Maximum 23.30 8.63 1.28 5.00 15.00 20.00 10.15 100.00 2.19 5.00 42.00 82.00 75 413.7 137.90 248.20 # 116.90 689.50 505.00

Count 36 36 36 36 36 36 36 36 36 36 36 36 3 15 15 15 15 15 15 540
Table 5 The matrix of Pearson correlation coefficients.

Octhedr Bulk
MDD OMC D10 D30 D50 D60 CGP Cu Cc Fines Sand Gravel RAP sig1 sig3 sigd J2 al Stress Stress MR
MDD 1
-
OMC 0.68468 1
0.12033
D10 7 -0.3312 1
0.39052 0.88478
D30 6 -0.47134 8 1
0.61677 0.45929 0.79541
D50 5 -0.46293 5 4 1
0.62744 0.32718 0.67890 0.98147
D60 9 -0.40883 8 5 5 1
0.60244 0.51621 0.82562 0.99706
CGP 5 -0.46126 6 5 7 0.9746 1
0.31786 0.06811 - - 0.37251 0.53549 0.37144
Cu 2 3 0.25315 0.10005 4 3 7 1
0.12900 0.77866 0.88727 0.63604 0.50650 -
Cc 3 -0.1575 9 7 5 6 0.65983 0.23461 1
0.47202 - - 0.34483 0.41929 0.32143 0.44747 -
Fines 2 -0.16106 0.21468 0.00655 6 4 9 2 0.24597 1
-
- 0.57363 - - - - - 0.2401
Sand 0.50084 2 0.72569 0.82883 -0.8183 0.78503 -0.85106 0.30518 0.51618 6 1
-
- 0.83171 0.60750 0.19071 0.09069 0.24931 - 0.39541 0.3704 -
Gravel 0.01112 -0.46567 4 1 3 8 8 0.24813 2 6 0.65351 1
- 0.16105 0.00655 - - - 0.24596 0.24015 0.37046
RAP 0.47202 5 0.21468 2 0.34484 0.41929 -0.32144 0.44747 8 -1 5 5 1
1.73E-
sig1 1.82E-17 5E-18 0 8.18E-18 -1.2E-17 1.94E-17 -5.1E-19 3.98E-18 1.03E-17 17 -1.1E-17 -4.3E-18 -2.2E-17 1
1.92E- 0.85027
sig3 0 1.04E-17 9.57E-18 -1.8E-17 7.53E-18 2.31E-17 9.18E-18 4.16E-17 2.34E-17 17 1.29E-18 1.01E-18 6.71E-17 6 1
0.94568 0.63298
sigd 9.17E-18 8.16E-18 -5.3E-18 1.91E-18 -1.6E-17 2.23E-17 1.26E-17 7.96E-18 -3E-18 0 -8E-18 -3.5E-18 -1.8E-17 2 1 1
-2.4E- 0.93172 0.94284 0.78807
J2 9.53E-18 5.09E-18 5.2E-19 -2.6E-18 -4.3E-18 7.13E-18 1.05E-17 2.65E-17 -1.1E-17 17 0 -4.1E-18 -3.4E-17 3 2 4 1
Octahedral -3.4E- 0.94568 0.63298 0.78807
Stress 1.34E-17 -1.7E-17 6.98E-18 1.69E-18 1.22E-18 5.92E-18 5.8E-18 4.61E-18 6.17E-18 18 9.44E-19 1.78E-17 1.73E-17 2 1 1 4 1
5.67E- 0.96831 0.83452 0.97366
Bulk Stress 3.96E-18 9.88E-18 2.6E-18 4.84E-18 2.78E-18 1.72E-17 1.74E-18 -6.3E-18 7.65E-19 18 -1.3E-17 -2.4E-18 2.18E-17 3 0.95478 7 7 0.834527 1
-
- 0.06922 0.15556 0.06798 - - 0.14702 0.2979 0.06468 0.29799 0.82025 0.86415 0.67273 0.84447 0.87336
MR 0.17586 7 2 5 0.11507 -0.1634 -0.10439 0.26007 5 9 6 0.15712 3 7 4 3 2 0.672733 7 1
Table 6. ANOVA result for the determination of statistically significant effects of input variables
Coefficient Standard
s Error t Stat P-value
Intercept -16.2602 19.03159 -0.85438 0.393447
OMC 12.08958 2.370768 5.099437 5.45E-07
Cu -0.85196 0.098533 -8.64647 1.63E-16
CGP 2.233116 1.008618 2.214037 0.027435
RAP 0.693034 0.085541 8.101777 7.96E-15
Gravel 0.769207 0.208472 3.689737 0.000258
sig3 1.466174 0.043895 33.40193 9.8E-114
sigd 0.23634 0.030394 7.775909 7.51E-14

Table 7 The best fit nonlinear functions for the individual input variables and MR

Input
Variables Rule R-Squared Equation
OMC Power 0.005 y = 143.71x0.2281
CGP Exponential 0.013 y = 242.48e-0.022x
Gravel Linear 0.026 y = 1.4246x + 208.27
Cu Exponential 0.072 y = 257.19e-0.005x
RAP Linear 0.088 y = 1.0967x + 177.48
Sigma-d Power 0.544 y = 37.645x0.4067
Sigma-3 Power 0.772 y = 32.044x0.4761
Where X is the input variable and y is the output variable (MR)

Table 8
Sr. Value used for this
Hypermarameter Description
No. study
The function used to measure the quality of a
split. For classification, options typically
include "gini" or "entropy", while regression MSE
Criterion for
1 can use "mse" (mean squared error) or "mae"
Splitting
(mean absolute error). The choice of criterion
can affect the model’s accuracy depending on
the type of data and the problem.
This can help control the growth of the trees, None i.e., there is
preventing overfitting by reducing no limit on the
complexity. number of leaf
Maximum Leaf
2 nodes. The tree will
Nodes
expand until it
meets other
stopping criteria.
A node will be split if the decrease in impurity 0.0 i.e., no
is greater than or equal to this value. his is minimum threshold
another way to control tree growth, helping to is enforced, and the
Minimum stop splits that do not add much value to the tree can grow as
3 Impurity prediction. long as other
Decrease stopping criteria
(like
min_samples_split)
are satisfied.
Controls the randomness of the bootstrapping 42
of samples and feature selection. Setting this
4 Random State parameter can make results reproducible,
allowing for consistent results on the same
data across runs
Larger leaf sizes can help smooth the model, 1 i.e., the minimum
reducing its sensitivity to noisy data. number of samples
required to be at a
Minimum
leaf node. Smaller
5 Samples per
values allow trees to
Leaf Node
grow deeper,
potentially
overfitting.
This parametere defines the model 1.0 i.e., which
complexity. Lowering max features reduces means all features
model complexity, often leading to less are considered for
Maximum
6 correlation between individual trees and a split
Features
more generalized model. Typical values are
"auto" (square root of the total number of
features), "sqrt", "log2", or an integer.
7 Bootstrap A boolean indicating if bootstrap samples True
Sampling (sampling with replacement) should be used
to grow trees. If set to True, each tree is
trained on a random sample of the data with
replacement, improving the ensemble’s
 generalization ability.
More trees can improve performance by
Number of trees reducing variance, but it also increases
8 Varied from 1 to 15
(n_estimators) computational cost. Too many trees may lead
to diminishing returns on accuracy.
Maximum depth of trees limits how deep each None i.e., the trees
tree can grow. Setting a maximum depth helps grow until all leaves
to control overfitting. Shallow trees may are pure (only one
Maximum underfit, while very deep trees can overfit, class or value) or
9
Depth of Trees capturing noise in the data. until all leaves
contain fewer than
min_samples_split
samples.
Maximum depth of trees limits how deep each
tree can grow. Setting a maximum depth helps
Mximum Depth
10 to control overfitting. Shallow trees may Varied from 1 to 15
of Tree
underfit, while very deep trees can overfit,
capturing noise in the data.
This parameter specify the strategy used to
choose the split at each node. Using "random"
11 Splitter can make the model faster and more Best
generalized, especially on large datasets,
though "best" is often more accurate.
This function specify the minimum number of
samples required to split an internal node.
Min Sample
12 Higher values prevent splitting nodes with 2
Split
few samples, helping to avoid creating overly
specific patterns (overfitting).
The minimum weighted fraction of the sum
Min Weight
13 total of weights (all samples) required to be at 0.0
Fraction Leaf
a leaf node.
Cost The complexity parameter used for Minimal
14 Complexity Cost-Complexity Pruning. Higher values 0.0
Pruning Alpha result in simpler (pruned) trees.
Note: parameter 1-6 are relevent to both RF and DT; 7-9 are relent to RF only; 10-14 are
relevent to DT only

Table 9
Sr. Value used for this
Hypermarameter Description
No. study
Defines the depth of the network. More
layers allow the model to learn more complex
1 to 3
Number of patterns but may increase the risk of
1
Hidden Layers overfitting. ypical Range: 1–5 layers
(depending on the complexity of the
problem).
Defines the width of each layer. A higher
Number of
number of neurons increases the model’s
2 Neurons per 1 to 16
capacity but also computational cost and the
Layer
risk of overfitting
Determines the non-linearity introduced at
each neuron. Common Choices:

 relu: Rectified Linear Unit (default for

Transfer most tasks).
3 Activation  sigmoid: For binary classification or Relu
Function probabilistic outputs.
 tanh: For bounded outputs in regression.
 softmax: For multi-class classification at
the output layer.

Determines the type of output the network

produces:
Output
 softmax: Multi-class classification.
4 Activation Linear
 sigmoid: Binary classification.
Function
 Linear: Regression.

Controls the step size during optimization.

Learning Rate Smaller values provide stable convergence but
5 0.01
Initial slower training. Typical Range: 0.001–0.01
(with adaptive optimizers like Adam).
Determines the algorithm used to update
weights. Common Choices:

6 Optimizer  SGD (Stochastic Gradient Descent) Adam

 Adam (Adaptive Moment Estimation)
 RMSprop (Root Mean Square Propagation)

The number of iterations in a Multilayer

Perceptron (MLP) refers to the number of
Maximum times the model updates its weights during
7 3500
Iterations training. This number depends on the batch
size, number of epochs, and the size of the
dataset.
8 Random State Controls the randomness of the bootstrapping 42
of samples and feature selection. Setting this
 parameter can make results reproducible,
allowing for consistent results on the same
data across runs
Specifiies the number of samples per batch for
9 Batch Size Auto
stochastic optimizers
Specifies the schedule for adjusting the
learning rate.
 ‘Constant’: initially specified learning rate
remaines constant
 'invscaling': Gradually reduces the learning
10 Learning Rate constant
rate.
 'adaptive': Keeps the learning rate constant
unless the training loss doesn't improve for
two consecutive epochs.