TAYLOR ROD IMPACT TEST
Figure-1-Schematic representation of Taylor rod test
Introduction
The Taylor rod impact test is a simple and widely used experimental method used to study the characteristics
of a material under dynamic load conditions in which a flat-ended cylindrical rod is propelled at high speed to
impact a rigid target, producing large plastic deformation at the impact end. By analysing the final deformed
shape of the rod and the extent of mushrooming and the length of the undeformed region-the test enables
researchers to characterize the dynamic yield strength and plastic behaviour of materials under impact loading
and extremely high strain rates.
Objective
From data points collected from simulation and literature train validate and test an Artificial Neural Network
(ANN) to predict the characteristics of the specimen after the impact.
Geometry and Material properties
For the simulation quarter model of a rod is constructed with initial radius 𝑅𝑖 = 3.7975mm and for various
length and velocity the impact simulation is done in Abaqus in explicit scheme using Johnson Cook material
model.
The parameter table and corresponding mesh size is tabulated according to[1].
L = 37.97mm
Fig2- Master surface V = 181
Table-1. Mesh discretization details of rod with 𝑟𝑖 = 3.795mm for all cases by[1]
Initial Length (𝐿𝑖)(mm) Initial Velocity(𝑉𝑖)(m/s) Mesh Discretization
11.39 285 62 x 41
15.19 234
275 161 x 41
302
25.29 170
268 x 41
215
37.97 181
183
224 401 x 41
234
270
56.96 242 601 x 41
Johnson-Cook Constitutive Model Equation:
𝜀 𝑇 − 𝑇room 𝑚
𝜎= (𝑎 )
+ 𝑏𝜀 (1 + 𝑐 ln ( )) (1 − (
𝑛
) )
𝜀0 𝑇𝑚 − 𝑇room
Table 2. Table of Material Properties
ρ(Density) 7830 kg/m³
E (Young’s Modulus) 82.9 GPa
K (Bulk Modulus) 169.1 GPa
c (Specific Heat Capacity) 460 J/kg·K
a (Yield Stress) 792 MPa
b (Hardening Modulus) 510 MPa
n (Strain Hardening Exponent) 0.26
c (Starin Rate Sensitivity) 0.014
m (Thermal softening Exponent) 1.03
Tₘ (Melting Temperature) 1293 K
𝑇𝑟𝑜𝑜𝑚 (Room Temperature) 293 K
ε₀̇ (Reference Strain Rate) 1 s⁻¹
Boundary Conditions
The boundary conditions assigned for the simulation is as follows,
1. ENCASTRE-On the rigid plate defining the structure as a rigid or fixed support, all the degrees of
freedom are restrained, preventing any rotation at that location.
Fig3-Rigid plate
2. XSYMM- On ADFC face as x-y plane is the contact plane, this makes a symmetric nodal condition
on that face so that Poisson's effect along that plane is neglected and ensuring no deformation occurs
normal to the plane.
Fig 4 – XSYMM surface
3. YSYMM- On ABED face as x-y plane is the contact plane, this makes a symmetric nodal conditionon
that face so that Poisson's effect along that plane is neglected and ensuring no deformation occurs
normal to the plane.
Fig 5– YSYMM surface
Numerical Parameters
1) Solver used is Abacus explicit solver with 50 intervals in the time step.
2) Element type used for meshing
Fig 6 – Element Type
3) Contact Algorithm- Surface to Surface contact explicit algorithm is used with initial step for analysis and
kinematic contact method is used to avoid part penetration and a frictionless penalty condition is applied.
Fig 7-Master surface Fig 8- Slave surface
Fig 9 -Interaction Type
Representative Mesh
Mesh Discretization for L= 37.97mm
Figure 10-41 X 41 element along x and y
Figure 11- 401 element along z
Validations
1) For final undeformed length the length of rod where the nominal radial strain or the logarithmic radial
strain value is less than 0.2%.
𝑙
ε = ln ( )
𝑙0
Figure12-From [2] X- length of undeformed part
|𝑋𝑝𝑟𝑒𝑠𝑒𝑛𝑡 − 𝑋[1]|
Percentage error =
𝑋[1]
X – length of undeformed part L1 – final length of rod
Table 3. Comparison of present result with results of [1] for final undeformed part
Length(mm) Velocity Final length of Final length of undeformed Percentage error
(m/s) undeformed rod(mm) 𝑋[1] rod(mm)𝑋𝑝𝑟𝑒𝑠𝑒𝑛𝑡
[1] (From ABAQUS)
11.39 285 3.2 or 2.7 4 25
15.19 234 5.6 or 5.3 6.95 24
275 5.2 or 4.7 6.37 22.5
302 5.2 or 4.3 5.52 6
25.29 170 11.2 or 10.4 8.8 15
215 10 9.47 5.3
37.97 181 15.5 or 17 16.2 4.7
183 15.3 or 17 15.6 1.9
224 11.9 or 15.5 15.5 0
234 13.3 or 15 16.38 9.2
270 11.8 or 12.9 15.7 17.8
56.96 242 17.5 or 22 22 0
2) Validation with final length after impact values
Table 4. Comparison of present result with results of [1] for final length of the part
Length(mm) Velocity Final length of rod(mm) 𝐿[1] Final length of Percentage error
(m/s) [1] rod(mm) 𝐿𝑝𝑟𝑒𝑠𝑒𝑛𝑡
(From ABAQUS)
11.39 285 9.2 or 9.6 9.26 0.6
15.19 234 13.1 or 13.3 13.17 0.5
275 12.4 or 12.8 12.5 0.8
302 12 or 12.5 12 0
25.29 170 23.4 or 23.3 23.29 0.4
215 22.4 or 22.5 22.3 0.4
37.97 181 34.6 34.4 0.5
183 34.7 or 34.6 34.5 0.2
224 33 or 33.3 32.9 0.3
234 33.1 or 32.9 32.6 1.5
270 31.3 or 32 33.2 3.75
56.96 242 47.8 or 48.7 48.4 0.6
|𝐿𝑝𝑟𝑒𝑠𝑒𝑛𝑡 − 𝐿[1]|
Percentage error =
𝐿[1]
3) Validation of Diameter of the impact face
|𝐷𝑝𝑟𝑒𝑠𝑒𝑛𝑡 − 𝐷[1]|
Percentage error =
𝐷[1]
Table 4. Comparison of present result with results of [1] for final diameter of impact face
Length(mm) Velocity Diameter of the impact Diameter of the impact Percentage error
(m/s) 𝐷[1]face(mm)[1] face (mm)𝐷𝑝𝑟𝑒𝑠𝑒𝑛𝑡
(From ABAQUS)
11.39 285 10.9 or 10.3 11.3 3.6
15.19 234 10.2 or 10.1 10.69 4.8
275 11.2 or 10.7 11.7 4.4
302 12.2 or 11.1 12.49 2.3
25.29 170 9.2 or 9.5 9.72 2.3
215 10.3 or 10.4 10.86 4.4
37.97 181 9.5 or 10 10.12 1.2
183 9.8 or 10 10.2 2
224 10.5 or 11 11.23 2
234 10.6 or 11.2 11.678 4.26
270 12.1 or 11.8 12.55 3.72
56.96 242 11.6 or 11.8 12 1.69
Table 5: Data obtained from finite element result for initial training.
Input Parameters output parameters
Test Initial Initial Impact Final Total Final Final Undeformed
No. Diameter Length Speed Length(mm) Diameter(mm) Length(mm)
Abaqus Abaqus Abaqus
1 7.595 11.39 285 9.26 11.3 4
2 7.595 234 13.17 10.69 6.95
3 7.595 275 12.5 11.7 6.37
15.19
4 7.595 288.5 12.2718 12.09944 5.4508
5 7.595 302 12 12.49 5.52
6 7.595 170 23.29 9.72 8.8
7 7.595 184.5 22.2955 10.088 9.95597
8 7.595 192.5 22.7637 10.2873 9.7892
25.29
9 7.595 198.5 22.6231 10.44368 9.81801
10 7.595 206.75 22.4284 10.6612 9.97437
11 7.595 215 22.3 10.86 9.47
12 7.595 181 34.4 10.12 16.2
13 7.595 182 34.4362 10.153 15.856
14 7.595 183 34.5 10.2 15.6
15 7.595 37.96 203.5
16 7.595 224 32.9 11.23 15.5
17 7.595 234 32.6 11.678 16.38
18 7.595 270 33.2 12.55 15.7
19 7.595 56.96 242 48.4 12 22
Training a neural network with BR-BP and K fold cross validation
A dataset comprising 17 samples was used to train a neural network using 5-fold cross-validation. The
dataset was partitioned into five approximately equal subsets, w here in each iteration, one subset served as
the validation set while the remaining four were used for training. This process was repeated five times to
ensure that each data point was used once for validation and four times for training. Such an approach helps
mitigate overfitting, provides a more reliable estimate of model performance, and ensures that the model is
exposed to the entire dataset during the training-validation cycle.
Table 6: Details of training dataset and testing dataset used in k-fold cross validation. The indices refer to the
case number in Table 5.
Split Number Training dataset Number of Testing Dataset Number of Testing Fold for
Indices Training Indices data(𝑁𝑡𝑟𝑎𝑖𝑛) Testing
Data(𝑁𝑡𝑟𝑎𝑖𝑛) Dataset
Split-1 1,2,5,6,8,10,11,12 13 3,4,7,9 4 Fold-1
,14,16,17,18,19
Split-2 1,2,3,4,5,6,7,8,9,1 14 12,14,19 3 Fold-2
0,11,16,17,18
Split -3 2,3,4,5,7,8,9,12,1 13 1,6,10,11 4 Fold-3
4,16,17,18,19
Split-4 1,3,4,5,6,7,9,10,1 14 2,8,16 3 Fold-4
1,12,14,17,19
Split-5 1,2,3,4,6,7,8,9,10, 14 5,17,18 3 Fold-5
11,12,14,16,19
The Neural Network architecture contains 2 input neurons and 9 hidden neurons in the hidden layer and 2
output neurons and the model is trained for a maximum of 1500 epochs.
Figure 13-Graph showing the MSE value for different hidden layer neurons all trained for 1500 epochs
Split Epochs MSE Table 7: Training parameter
for best configuration (2-9-
Split-1 1500 0.000273 2)
Split-2 1500 0.022879
Split -3 1500 0.001252
Split-4 1500 0.000414
Split-5 1500 0.055483
After training and testing a random data point is checked
The two input values initial length and impact velocity: 37.97mm 203.5m/s
Predicted Output from Neural network: 31.9269mm 12.0107mm
Similarly a few data points were checked with the finite element simulations as well as through
neural network.
Table 7: Comparison table
Input Values Result obtained from Finite Result obtained from Percentage error(%)
element simulation Neural Network
Length Velocity Final Length Final Final Length Final 𝐿𝐹𝐸 − 𝐿𝑁𝑁 𝐷𝐹𝐸 − 𝐷𝑁𝑁
(mm)( 𝐿𝑖) (m/s) (mm)𝐿𝐹𝐸 Diameter (mm) 𝐿𝑁𝑁 Diameter 𝐿𝐹𝐸 𝐷𝐹𝐸
( 𝑉𝑖) (mm) 𝐷𝐹𝐸 (mm) 𝐿𝑁𝑁 × 100 × 100
37.97 182 34.4362 10.153 34.4605 10.1619 0.07 0.087
37.97 203.5 33.7 10.75 31.927 12.0107 5.26 11.73
31.63 189.5 28.612 10.29 28.503 9.848 0.3 4.29
31.63 198 28.3894 10.152 28.618 10.1286 0.8 0.23
31.63 206.5 28.1 10.767 28.6735 10.3786 2 3.6
47.965 211.5 41.827 10.99 44.2373 12.0698 5.7 9.8
From Table-7 we can say that the results obtained from the neural network are quite matching with
the finite element study.
Data Generation and Hyperparameter Tuning Date - 14/06/25
Around 73 simulations were performed but for Training and Development of the neural network only 57
data sets were used. The input data set were Initial length (𝐿𝑖) and Velocity ( 𝑉𝑖) and Output Data set Final
Length (𝐿𝐹) and Final Diameter (𝐷𝐹).
The number of neurons in the hidden units in the Neural Network is changed from 1-15 and the MSE and
R(Pearson correlation coefficient) is calculated for each hidden unit. The code is updated to run each neuron
configuration across 10 independent runs and report the average and standard deviation of both MSE and
Pearson R, to identify consistently good architecture.
Best Model (Stable): 12 neurons with Average MSE = 0.006528 and R = 0.92
Figure14- Graph showing the average MSE and R values for different hidden
layer neurons all trained for 1500 epochs
Neural network architecture of 2-12-2 was trained and validated with 5-fold cross validation and the
Training MSE and Validation MSE is being analysed to see the possible overfitting (high variance)
or under fitting (high bias) in each split 75% of data is taken as training set and 25% for validation.
Table8-Percentage error on validation and training set in each folds
FOLD Training set MSE Validation Set MSE Average % from Train Average% error from
set Validation set
Final Final Final Final
length diameter length diameter
1 0.002402 0.021114 1.29% 3.24% 1.27% 10.81%
2 0.005316 0.008054 0.83% 3.28% 1.51% 5.70%
3 0.006933 0.000896 1.69% 6.25% 1.11% 2.91%
4 0.006699 0.000770 1.00% 5.48% 1.49% 2.03%
5 0.007056 0.001139 1.51% 5.00% 3.03% 2.81%
From Table8-it can be inferred that the network is likely overfitting, as the validation set MSE is
nearly an order of magnitude higher than the training set MSE in some folds. This indicates that the
model may be capturing noise or specific patterns in the training data that do not generalize well to
unseen data.
Table9-Comparison between FEM results and Neural Network results
Input Values Result obtained from Finite Result obtained from Percentage error(%)
element simulation Neural Network
Length Velocity Final Length Final Final Length Final 𝐿𝐹𝐸 − 𝐿𝑁𝑁 𝐷𝐹𝐸 − 𝐷𝑁𝑁
(mm)( 𝐿𝑖) (m/s) ( (mm)𝐿𝐹𝐸 Diameter (mm) 𝐿𝑁𝑁 Diameter 𝐿𝐹𝐸 𝐷𝐹𝐸
𝑉𝑖) (mm) 𝐷𝐹𝐸 (mm) 𝐿𝑁𝑁 × 100 × 100
33.22 226 28.814 11.376 28.6968 10.8361 0.41 4.74
33.22 194.5 29.777 10.457 29.574 10.1817 0.682 2.63
33.22 187.25 29.9843 10.2645 29.7839 10.0717 0.668 1.88
33.22 222 28.9364 11.2543 28.8048 10.7359 0.455 4.6
40.35 208 36.4224 10.8864 35.30320 10.6257 3.1 2.39
40.35 202 35.9884 10.6936 35.4661 10.4883 1.45 1.92
50.25 239 42.5694 11.9136 42.8713 12.0206 0.71 0.8
From Table9 -the results obtained from Neural network is quite matching with the FEM results.
Data Generation and Hyperparameter Tuning Date - 21/06/25
Data sets from 108 simulations are taken for Training and Development of the neural network. The input
data set were Initial length (𝐿𝑖) and Velocity ( 𝑉𝑖) and Output Data set Final Length (𝐿𝐹) and Final Diameter
(𝐷𝐹).
A 3-layer feedforward neural network, consisting of 2 input neurons, a single hidden layer with a
variable number of neurons (1–10), and 2 output neurons, was trained using Bayesian Regularized
Backpropagation (BR-BP). To ensure robust generalization and mitigate overfitting due to random
weight initialization, each network configuration was independently trained 15 times using a fixed 5-
fold cross-validation scheme. The hidden layer activation function was set to either logsig (log-
sigmoid), tansig (hyperbolic tangent), and the activation function yielding the lowest average MSE was
selected as optimal and the number hidden layer neuron with the least MSE is found from the plot to
make final Neural Network architecture.
Figure15-Avg MSE vs Number of hidden layer neurons
for logarthmic sigmoid Activation function.
Best Number of Neurons: 5 (Minimum Avg MSE =
0.003089)
Figure16-Avg MSE vs Number of hidden layer neurons
for hyperbolic tangent Activation function.
From the Figure15 and Figure16 it can be noticed that logarithmic sigmoid activation function in the hidden
layer is giving minimum MSE throughout. Based on this result, the optimal neural network architecture is
selected as 2–5–2, comprising 2 input neurons, 5 hidden neurons using logsig activation, and 2 output neurons
with linear activation.
Table10-Parameters in 2-5-2 network
Fold Epochs Gradient Lambda MSE SSE SSW
1 84 1.35e-04 5.00e+10 0.000176157 0.0078 18.2779
2 125 1.63e-04 5.00e+10 0.0004272567 0.0188 18.3360
3 387 1.34e-04 5.00e+10 0.00283156 0.1246 15.8145
4 176 5.43e-06 5.00e+10 0.00399066 0.1676 823.7703
5 1008 5.95e-05 5.00e+10 0.00888812 0.3733 16.6190
Final 2-5-2 Model Summary:
Average MSE across 5 folds: 0.003262751
Total SSE = 63.724413
Sum of Squared Weights (SSW) =736.812338
Average epochs used: 356
Figure17 – Scatter plot of final deformed length for a given ( Li ) and ( Vi ) obtained from FEM
Figure18 – Scatter plot of final deformed length for a given ( Li ) and ( Vi ) obtained from NN
Figure19– Scatter plot of final diameter for a given ( Li ) and ( Vi ) obtained from FEM
Figure20 – Scatter plot of final diameter for a given ( Li ) and ( Vi ) obtained from NN