Page 83 - IJAMD-1-2

International Journal of AI for
Materials and Design
AMTransformer for process dynamics

approach, AM offers unparalleled capabilities for creating inflexible and unable to adapt to different datasets or AM
designs with advanced geometrical, hierarchical, material, methods. In addition, AM processes involve a complex
and functional complexities. These capabilities have hierarchy of interconnected points, lines, and layers.
significantly advanced innovation in both the design and Therefore, it is crucial to develop an ML method that
functionality of final products. 1-4 provides a comprehensive understanding of this entire
hierarchy, something that existing methods are insufficient
Among the various AM processes, laser-based methods,
such as laser powder bed fusion (LPBF) and direct to achieve. Consequently, there is an urgent need for
an alternative approach capable of encompassing the
energy deposition (DED), stand out. These methods use multi-scale spatial and temporal modeling required for a
a laser beam as an energy source to fuse powder-like raw
materials and build final products. In these processes, complete understanding of AM dynamics.
5,6
melt pools are generated when the laser interacts with To address this challenge, we first formulated the key
the raw materials. Melt pools are the areas where the laser dependencies in AM dynamics in this study. Building on
melts the powder materials, dynamically changing their this formulation, we proposed a novel method called the
states as the laser beam moves along the designated path. AMTransformer, designed to improve the understanding
7,8
The melt pools undergo continuous dynamic changes due of melt pool dynamics and predict its behaviors. The
to spatial and temporal dependencies between the physical AMTransformer leverages Koopman theory and a
entities involved in the AM processes. For example, the transformer architecture to capture spatiotemporal
high heat transfer within the melt pool results in spatial dependencies within and between melt pools at multiple
variations in temperature distribution. Furthermore, the scales, thereby providing a deeper understanding of the
thermal behavior of the melt pools changes over time due underlying dynamics that govern AM processes.
to the accumulation of layers and the presence of adjacent, The remainder of this paper is organized as follows: In
previously formed melt pools along the path. The physical Section 2, we review relevant literature on melt pool modeling,
9
dynamics of the melt pools during the process significantly a representative area of AM dynamics modeling. Section 3
impact the quality and properties of the manufactured establishes the problem statement, while Section 4 introduces
products. 10,11 the proposed method. In Section 5, we present a case study.
Machine learning (ML) can significantly enhance Finally, Section 6 includes a discussion, and Section 7 offers
our understanding of the complex dynamics involved concluding remarks and outlines future work.
in laser-based AM processes. By leveraging novel in situ
data generated during LPBF and DED operations (e.g., 2. Literature review
thermography, 13,14 visual imaging, 9,14-16 acoustic signals, 17,18 Deep learning methods have proven effective in identifying
and spectral data 19,20 ), ML algorithms can identify intricate and extracting significant spatial features from data, thereby
patterns and dependencies that traditional analytical or enabling precise predictions or classifications of melt pool
simulation methods might overlook. 21,22 For example, ML characteristics. Yang et al., used a convolutional neural
23
models can predict the spatial and temporal dependencies network (CNN) architecture, coupled with fully connected
within and between melt pools, accounting for factors like layers, to classify melt pool types using real-time melt
high heat transfer and the cumulative effects of successive pool monitoring (MPM) images. Similarly, Zhang et al.
24
layers. Enhancing the understanding of AM dynamics proposed a method that utilizes a hybrid of two CNNs to
9,10
through ML-driven real-time monitoring and predictive monitor and predict melt pools. Fathizadan et al. adopted a
14
analytics allows for adaptive control of the AM process, convolutional autoencoder model to derive a comprehensive
thereby improving consistency and reducing defects. This representation of melt pool areas, which they used to
data-driven approach not only enhances process efficiency distinguish anomalies through a clustering algorithm.
but also contributes to the development of more robust In addition, studies have explored the application of
and reliable AM systems capable of producing high- time-series neural networks to learn temporal dependencies
quality components with complex geometries and material involved in melt pool dynamics. Larsen and Hooper
25
processing. Advances in process monitoring technologies developed a method for predicting melt pool behavior
have further enabled the collection of novel in situ data by integrating a variational autoencoder with a recurrent
from AM processes, underscoring the importance of ML in neural network (RNN). In their approach, MPM images
comprehending and anticipating AM dynamics. 22
are first embedded into a latent space using a variational
However, existing methods using ML for analyzing autoencoder. This latent representation then serves as the
complex AM dynamics often lack generalizability. They input for a dynamics model, which employs an RNN to
rely on specific AM cases or data types, rendering them predict the distribution of the next state based on previous

Volume 1 Issue 2 (2024) 77 doi: 10.36922/ijamd.3919