Page 62 - IJAMD-1-1

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

3.1.3. Deep learning with augmented experiment data, a fine-tuning step was introduced. For the pre-training
results and fine-tuned simulation dataset of the DL-AugSim-Sim-Exp model, the data augmentation
In the third baseline model, DL-AugExp-sim consists of a technique was applied to simulation data, resulting in the
general neural network with two layers, each comprising generation of a total of 625 datasets, as illustrated in Figure 6.
16 nodes. This model was trained with 125 augmented Of these, 500 data points (80%) were used as training data,
experiment datasets and subsequently fine-tuned with 81 and the remaining 125 data points (20%) served as validation
simulation datasets. The model was subsequently tested with data. The final model was evaluated using experiment
ten experiment datasets. The augmented experiment data datasets. The validation dataset is used to provide an
encompasses input parameters such as LP, PMFR, and SS, impartial assessment of how well the model fits the training
while the simulation data include additional sulfur content data while tuning the model’s hyperparameters, thereby
input parameters. The augmentation procedure aligns with that preventing overfitting. Subsequently, the DL model was
used in the DL-AugExp model, incorporating additional data further fine-tuned with 81 simulation data points generated
of constant sulfur content at 0.0038 wt% to the augmentation through a full factorial design from Table 1. Additional
experiment dataset. This addition ensures a similar dataset fine-tuning was performed using 24 experimental datasets,
size during training. The total mean R is 0.8627, with a incorporating variations in experimental results to simulate a
2
corresponding standard deviation of 0.02362 (Table 7). real-world scenario, as depicted in Figure 17. The total mean
R is calculated to be 0.8943, with a corresponding standard
2
3.1.4. Deep learning with augmented simulation deviation of 0.03655 (Table 7).
dataset
3.2. Deep learning model trained with augmented
In the fifth baseline model, DL-AugSim consists of a general experiment data and fine-tuned with simulation
neural network with two layers, each comprising 16 nodes. and experiment data
This model was trained with 625 augmented simulation
datasets and was subsequently tested with 10 experiment The DL-AugExp-sim-exp consists of a general neural
datasets. The augmented simulation dataset encompasses network with two layers, each comprising 16 nodes. This
input parameters such as LP, SS, PMFR, and sulfur content. model was trained with 125 augmented experiment datasets
The augmentation procedure aligns with the process depicted and subsequently fine-tuned with 81 simulation datasets
in Figure 6 and follows the discretized levels outlined in (Figure 18). An additional step was introduced, using
Table 5. The total mean R is calculated to be 0.8565, with a 24 experiment datasets to further fine-tune the model.
2
corresponding standard deviation of 0.02342 (Table 7). Subsequently, the model was tested with 10 experiment
datasets. The augmented experiment data encompass
3.1.5. Deep learning with augmented simulation data input parameters such as LP, PMFR, and SS, while the
and fine-tuned with simulation dataset simulation data incorporate additional sulfur content input
In the last baseline model, DL-AugSim-sim consists of a parameters. The augmentation procedure mirrors that used
general neural network with two layers, each comprising in the DL-AugExp model, with the additional constant
16 nodes. This model was trained with 625 augmented sulfur content data at 0.0038 wt% to the augmentation
simulation datasets and further fine-tuned with 81 experiment dataset, ensuring a consistent dataset size
2
simulation datasets. Subsequently, the model was tested with during training. The total mean R is calculated to be
10 experiment datasets. The augmented simulation dataset 0.9086, with a corresponding standard deviation of 0.02818
encompasses input parameters such as LP, SS, PMFR, and (Table 7). The results produced were compared between
sulfur content. The augmentation procedure aligns with the the DL prediction model and ground truth experimental
process depicted in Figure 6 and follows the discretized levels results (Figure 19), where the Y-coordinate value (Y value)
2
outlined in Table 5. The fine-tuning process involves training would be used to calculate the R and RMSE.
the model with a smaller and more focused simulation The integration of simulation datasets into the training
dataset. The total mean R is calculated to be 0.8103, with a of deep learning models facilitates the prediction of
2
corresponding standard deviation of 0.04791 (Table 7). variations in sulfur content on geometry by the models.
Input parameters for the DL-AugExp-Sim-Exp model were
3.1.6. Deep learning with augmented simulation configured with varying sulfur content levels, specifically at
data and fine-tuned with simulation dataset and 0.003 wt%, 0.015 wt%, and0.03 wt%. These parameters were
experiment dataset employed to predict the final melt pool geometry. The results
To minimize the risk of overfitting during the training of deep reveal a correlation between the sulfur content and the
learning models on augmented experiments and simulation increase in the melt pool depth. This observation aligns with

Volume 1 Issue 1 (2024) 56 https://doi.org/10.36922/ijamd.2355