Page 61 - IJAMD-1-1

International Journal of AI
for Material and Design Integrating physics data for DL in DED

to establish a performance benchmark, employing the 16 nodes. This model was trained with 125 augmented
TensorFlow library. 38,39 Each model adhered to the training experiment data points and subsequently tested with 10
data listed in Table 6 and was tested on ten experiment experiment data points. The experiment data include input
results unseen by the ML model. The nomenclature and parameters such as LP, PMFR, and SS. The augmentation
characteristics of the deep learning model are detailed below: procedure, illustrated in Figure 6, involves curve-fitting

• The model “DL-Exp” represents the deep learning a polynomial regression model to the experiment data
model trained on experiment data. from Table 2. Subsequently, the discretized parameters,
• “DL-AugExp” represents the deep learning model as outlined in Table 8, were used to generate the dataset
trained with augmented experiment data. for training DL-AugExp. The total mean R is calculated
2
• “DL-AugExp-Sim” represents the deep learning to be 0.9362, with a corresponding standard deviation of
model trained with augmented experiment data, 0.02221 (Table 7).
incorporating fine-tuning using simulation data.
• “DL-AugExp-Sim-Exp” represents the deep learning
model trained with augmented experiment data, Table 6. Model and dataset set up for each case
undergoing fine-tuning using simulation data and Model name Training dataset type Shape of
subsequent fine-tuning with experiment data. training data
• “DL-AugSim” represents the deep learning model DL-Exp Experiment (24,3)
trained on augmented simulation data. DL-AugExp Augmented experiment (125,3)
• “DL-AugSim-Sim” represents the deep learning DL-AugExp-Sim Augmented experiment, (125,4)
model trained with augmented simulation data, simulation
incorporating fine-tuning using simulation data. DL-AugExp-Sim-Exp Augmented experiment, (125,4)
• “DL-AugSim-Sim-Exp” represents the deep learning simulation, experiment
model trained with augmented simulation data, DL-AugSim Augmented simulation (625,4)
undergoing fine-tuning using simulation data and
subsequent fine-tuning with experiment data. DL-AugSim-Sim Augmented simulation, (706,4)
simulation
Models exclusively trained on experiment data, such as DL-AugSim-Sim-Exp Augmented simulation, (730,4)
“DL-Exp” and “DL-AugExp,” feature three input features. simulation, experiment
Given that experiment data does not encompass sulfur
content as an input parameter, the dimensions of the
2
feature input take shapes of (24,3) and (125,3), respectively Table 7. Mean R , standard deviation (SD), and root mean
(Table 6). Each model underwent 100 runs with different square error (RMSE) of deep learning models
random states for testing and training the ML model, Model name Mean total R 2 SD total R 2 Mean total
ensuring the reliability of the predicted results. RMSE
DL-Exp 0.9015 0.05795 140.654
3.1. Baseline models
DL-AugExp 0.9362 0.02221 114.589
3.1.1. Deep learning with experiment results DL-AugExp-sim 0.8627 0.02363 161.518
In the first baseline model, DL-Exp consists of a general DL-AugExp-Sim-Exp 0.9086 0.02818 136.744
neural network with one layer consisting of three input DL-AugSim 0.8565 0.02342 165.768
nodes, two layers containing 16 hidden nodes, and six DL-AugSim-Sim 0.8103 0.04792 188.998
layers of output nodes. This model was trained with 24 DL-AugSim-Sim-Exp 0.8943 0.03655 146.552
experiment data points and was subsequently tested with
10 additional experiment data points. The experiment
data encompass input parameters such as LP, PMFR, and Table 8. Discretized parameter level for augmentation of
SS. The total mean R calculated for these testing data in experiment data
2
comparison with the prediction model is 0.9015, with a
corresponding standard deviation of 0.05795 (Table 7). Input parameters Levels
LP (W) 1000 1150 1300 1450 1600
3.1.2. Deep learning with augmented experiment SS (mm/min) 1000 1150 1300 1450 1600
results
PMFR (g/min) 14 15 16 17 18
In the second baseline model, DL-AugExp consists of a Abbreviations: LP: Laser power, PMFR: Powder mass flow rate,
general neural network with two layers, each containing SS: Scanning speed.

Volume 1 Issue 1 (2024) 55 https://doi.org/10.36922/ijamd.2355