Page 53 - IJAMD-1-1

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

The highlight of this study lies in the integration were conducted using central composite design (CCD)
of in-process data, which is unattainable through and curve fitting with a response surface methodology
experimentation. Pre-process and post-process data (RSM) regression model, yielding 20 samples of results.
obtained experimentally encompass laser power (LP), The parameter ranges were set as follows: LP ranged
powder mass flow rate (PMFR), scanning speed (SS), from 1000 W to 1600W, PMFR varied between 14 g/min
and melt pool contour. These sets of data will be referred and 18 g/min, and SS spanned from 1000 nm/min to
to as experimental datasets in the subsequent sections. 1600 mm/min. For constructing the CCD, a three-level
Conversely, the in-process and post-process data obtained design (-1, 0, +1) was employed, resulting in a 2-level,
from physics-based simulation include LP, PMFR, SS, sulfur 3-factor full factorial design with repeated center
content, and melt pool contour, collectively known as the points. Axial points were calculated using the following
simulation dataset in the subsequent sections. The sulfur formula:
content is critical data that prove challenging to obtain max min−
experimentally. Within this framework, it plays a pivotal role Axial points = ± α * + centerpoint (I)
in facilitating the integration of physics and experiment data, 2
enabling in-depth exploration of sulfur content variations in where α set to 1.633, establishing desired axial positions
relation to melt pool geometry. The integrated framework, at LPs of 810 W and 1790 W, PMFRs of 12 g/min and
as depicted in Figure 2, serves as a visual representation of 19 g/min, and SSs of 810 mm/min and 1790 mm/min.
the process of combining pre-, in-, and post-process data. The design matrix was implemented in a randomized
sequence to minimize potential systematic errors, as
2. Methods presented in Table 2. To broaden the experimental dataset
2.1. Generation of melt pool contour data from and encompass a wider parameter range, eight datasets
numerical simulation from full factorial design experiments within the CCD
range and eight datasets from outside the CCD range were
The melt pool simulation was used to generate SS316L single- conducted, as presented in Table 3 and Table 4. For cross-
track deposition data for training the ML model (Figure 3). sectional samples, the printed single track was cut using a
This simulation utilized the temperature gradient of surface Makino U3 Wire Electrical Discharge Machine (WEDM,
tension within the Fe-S system, a derivation elucidated Makino Asia, Singapore), and track height, width, and
in references. 30,31 Rigorous validation of the melt pool depth were measured using Keyence VHX-7000 (Keyence
simulation was achieved through the comparisons of single Corporation, Japan). Contour points along the deposition
tracks printed experimentally with known sulfur content.
and dilution contours were measured, as depicted in
The simulation of the single-track melt pool involved Figure 5. These contour points from both experiment
the application of pre-determined process parameters and simulation data were used to curve-fit a polynomial
(Figure 3A). Identification of the track depth occurred equation, yielding bead coefficients, namely “a0,” “a1,”
within the cells in the substrate where the temperature and “a2,” and dilution coefficient, namely “b0,” “b1,” and
surpassed the melting point of the material. Subsequently, “b2,” which function as the output data in this study. It is
the final numerical result underwent slicing at the midpoint essential to note that, in this study, direct measurements
of the single-track melt pool simulation, yielding the cross- of maximum track height, width, and depth were not
section of the melt pool (Figure 3B). From these results, the conducted. Instead, information on track height, depth,
contour coordinates in the (X, Y, Z) format were extracted, and width along the perimeter of the deposited track was
constituting a list of points (Figure 3C). captured through contour points along the deposition and
The contour data were generated by manipulating dilution contours.
parameters including LP, PMFR, SS, and varying levels of 2.3. Data pre-processing and evaluation
sulfur content. These variations aimed to study the effect
of sulfur content on the final track dimensions (Table 1). 2.3.1. Data pre-processing
The sulfur content was deliberately selected at 0.0005 Due to limited resources, this study incorporated only a
wt%, 0.004 wt%, and 0.015 wt% due to the non-linear small set of experimental data. Recognizing the substantial
relationship between sulfur content and the temperature data requirement for deep learning models, the number
gradient of surface tension (Figure 4). 30,31 of data points obtained from either experimental mean or
numerical simulation alone is insufficient. Consequently,
2.2. Experiment setup and results the data augmentation technique proposed by Chen et al.
32
A full factorial design comprising 3 cases was simulated, was employed to expand the database. Depending on the
4
resulting in a total of 81 cases. Single-track experiments dataset used for training, the RSM regression model was

Volume 1 Issue 1 (2024) 47 https://doi.org/10.36922/ijamd.2355