Page 51 - IJAMD-1-1

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

trial-and-error approach. This challenge becomes more Mozaffar et al. employed a recurrent neural network
28
pronounced when the process is repeated with a set of with a gated recurrent unit technique to predict thermal
materials featuring slight changes in critical elements. In histories, accounting for changes in geometry and process
addition, experimentation alone proves insufficient in parameters. To address the challenge posed by varying
capturing all influencing factors of a print arising from a geometries in AM processes, a graph-based representation
material change. One example is the impact of surface- was further developed using neural networks to capture
29
active elements, such as sulfur, wherein the sulfur content spatiotemporal dependencies of thermal responses.
21
within the deposition of stainless steel influences the Lu et al. used an improved backpropagation-based
track geometry of a single-track print in directed energy network to predict the height of the single track. Wang
19
deposition (DED). 1,2 et al. developed a Gaussian process regression (GPR)
model using physics-based simulation data to predict
Recognizing the advantages of physics-based models,
which possess the capability to reveal underlying features the geometrical characteristics of cladding tracks. It is
noteworthy that the GPR model was exclusively trained
unattainable through experimental means, numerous using physics-based data, potentially introducing inherent
attempts have been made to model the DED process biases due to missing physics. Furthermore, these models
using physics-based simulation. Key phenomena, were used to analyze process and output relations that can
3
such as the effect of sulfur content, have been captured be experimentally measured but do not fully capture the
using computational fluid dynamics coupled with advantages of physics-based simulations. However, ML
interface tracking models such as volume of fluid. 2,4-6 methods, characterized as black-box methods lacking
Gan et al. investigated the influence of sulfur content knowledge of physics, necessitate meticulous training
6
and temperature on Marangoni convection using an using available experimental data – often fraught with
improved surface tension model at the top of the melt inevitable measurement errors. This limitation impinges
pool. Zhang et al. studied the effect of sulfur content on upon the interpretability, applicability, generalizability,
2
track geometry, discovering that an increase in sulfur and transferability of ML methods in broader process
content correlates with an increase in melt pool depth. conditions. 7
These studies strongly underscore the advantages of
physics-based models, revealing underlying phenomena Driven by the respective advantages and disadvantages
that cannot be captured experimentally. However, physics- inherent in physics-based simulation and ML, this
based models are innately biased due to assumptions paper advocates for the integration of physics-based and
made to reduce the complexity of modeling the process. experiment-based datasets using ML. In addition, the
7
Furthermore, numerical models demand the calibration extensive exploration of integrating critical in-process
of their parameters and prove too resource-intensive for information, such as sulfur content, with pre- and post-
real-time forecasting. While most physics-based models process information through physics-based simulation
7
accept process parameters as input to obtain output, such has been lacking. Consequently, the objective of this
as fusion zone geometry, temperature fields, cooling rates, paper is to develop an ML model leveraging data collected
and temperature gradients, their practical use in shop floor throughout the DED process, classifying it into pre-process,
settings is limited owing to their nature as forward models. in-process, and post-process data. Physics-based melt pool
Users often seek parameters tailored to a specific cooling simulation data were generated to capture the in-process
rate or fusion zone geometry; however, the forward model dataset, which includes variations of sulfur content on final
geometry. Simultaneously, an experiment was conducted
cannot identify these process variables without undergoing to obtain data that captured the pre-process and post-
multiple attempts and adjustments. 8
process datasets, which included variations of process
Machine learning (ML) has proven to be an effective parameters on final geometry. The dataset is integrated
alternative for addressing complex problems. 9-11 The using a deep learning model with a specific training
advantage of ML includes the ability to establish sequence. The best-performing deep learning model for
relationships between process parameters and final track geometry prediction, denoted as DL-AugExp-Sim-
printed characteristics, encompassing microstructure, 12-14 Exp, was developed through rigorous comparisons with
defects such as pores or cracks, 7,15-18 geometrical the performances of six other baseline models.
characteristics, 11,17,19-22 and mechanical properties such as
hardness, tensile strength, and fatigue. 23-25 Furthermore, 1.1. Integration of pre-process, in-process, and post-
ML exhibits the capability to manage large, diverse, and process information
complex datasets, facilitating the extraction of useful In the DED process, the information encompassing
real-time information with ease. 8,26,27 In DED processes, the entire procedure can be classified into pre-, in-, and

Volume 1 Issue 1 (2024) 45 https://doi.org/10.36922/ijamd.2355