Page 81 - ESAM-1-1

Engineering Science in
Additive Manufacturing ML in additive manufacturing

physics-informed learning) being developed. Zhang et al. (3) Structure ViTs, GPTs, and BERTs. To learn AM
12
reviewed AM datasets to identify their size and found only structure spaces for effective process-structure and
a few applications exceeding 10,000 data instances for ML structure-property linkages as well as model structure
as shown in Figure 10. This highlights the need to develop representations for efficient analysis and post-process
large datasets by standard compilation and processing to performance prediction.
leverage the strengths of advanced learning algorithms.
6.4. Resource efficiency and scalability
6.3. AM foundation models The growing adoption of ML applications in AM requires

With hundreds of existing research efforts at the initiatives and practices to support efficient and scalable AI
intersection of AM and ML, AM technology can benefit development. Scalable AI development refers to designing,
from specialized foundation models to address existing building, and deploying ML models that can handle
challenges in design, process, and structure control. increasing amounts of data, computational burden, and
Foundation models are large-scale DL architectures complex AM concerns without significant degradation in
that can model and generate insights relevant to specific performance. Efficient AI refers to ML-based systems that
fields. In science and engineering discipline, numerous achieve high performance while minimizing the use of
135
efforts have focused on domain-specific vision and resources (e.g., time, computation, and cost). Collectively,
language modalities to develop foundation models by these techniques can enhance the feasibility and effectiveness
customizing transformer architectures such as generative of ML applications in AM by optimizing model development.
pre-trained transformers and bidirectional encoder Below are a few research directions linked to the efficient
representation from transformers or BERT (e.g., SciBERT and effective applications of ML techniques in AM:
136
for scientific information extraction, SegGPT for structure (1) Parameter-efficient training. To achieve high model
semantic segmentation, and MedViT for medical image performance through efficient learning practices such
85
classification ). The development opportunities for AM as knowledge transfer and adaptation of pre-trained
137
foundation models are categorized below: models, few-shot learning, and meta-learning. 138
(1) Design ViTs, GPTs, and BERTs. To encode AM design (2) Quantum ML. To effectively model complex process
knowledge and practices through LLMs and to learn representations in AM by leveraging quantum
design spaces through large 3D vision models computing during the training process. In this regard,
(2) Process ViTs, GPTs, and BERTs. To model process microstructure modeling, process optimization, and
phenomena through common process emissions and quantum-enhanced simulations are the most relevant
signals as well as to support process optimization and candidates for future research. 139
quality control (3) Model compression and pruning. To reduce the
size of large models without significantly affecting
their performance. This is usually accomplished
by eliminating unnecessary or less useful model
parameters. Another approach of model pruning is
quantization which reduces the precision of model
parameter values. 140
(4) Edge AI and federated learning. To enhance the
scalability and efficiency of ML models in AM,
computation can be distributed across local devices
such as sensors and controllers. Training models
collaboratively on decentralized data sources reduces
reliance on a central server, limiting data transfer and
improving privacy. This approach maintains model
performance while lowering computational demands
and supporting real-time decision-making in diverse
manufacturing environments. 141-143
Figure 10. Reviewed AM dataset sizes in 2021 for ML applications. There
is a need to expand AM dataset size to effectively leverage large-scale ML 7. Challenges and opportunities in
models. 15.3% of the reviewed articles did not specify the dataset size,
highlighting the need to provide reproducibility critical information in industrial deployability
AI-driven AM.
Abbreviations: AI: Artificial intelligence; AM: Additive manufacturing; Recent advancements in AI-driven AM have predominantly
ML: Machine learning. concentrated on training advanced and accurate models,

Volume 1 Issue 1 (2025) 14 doi: 10.36922/ESAM025040004