Page 74 - ESAM-1-1

Engineering Science in
Additive Manufacturing ML in additive manufacturing

Figure 6. Emerging learning strategies in additive manufacturing,
2015 – 2025. Physics-informed learning encompasses all techniques
Figure 5. ML strategies in AM applications. The learning strategies in the that integrate domain knowledge (e.g., physical laws, process physics,
past 11 years have been analyzed from 1250 research articles (excluding context awareness, application constraints, and prior knowledge) into the
conference papers and book chapters) as follows: 3 (2015), 4 (2016), learning process.
7 (2017), 34 (2018), 54 (2019), 88 (2020), 135 (2021), 215 (2022),
245 (2023), 414 (2024), and 51 (2025).
Abbreviations: AM: Additive manufacturing; ML: Machine learning. size, dimension, and modality. While the initial applications
leveraged simpler linear models (e.g., linear, polynomial,
guided approaches to tailor ML algorithms and pipelines ridge, or lasso regression ) for classification and regression
47
and enhance their effectiveness to model AM concerns. tasks linked to different process concerns (e.g., parameter
5
Knowledge transfer strategies are being leveraged to prediction and state or regime classification), recent
expedite the development of ML models in newer AM research has applied more advanced ML architectures. In
contexts (e.g., materials, machines, processes, and quality 2018, CNNs 17,73,74 and recurrent neural networks (RNNs)
75
concerns) by transferring knowledge from existing data- were applied for the 1 time, highlighting an extension
st
driven solutions. 66-68 In the beginning, knowledge transfer from FFNNs to handle time-series and image datasets
76
approaches in AM leveraged generic real-world datasets from AM processes. In 2020, recurrent CNNs (RCNNs)
77
and pre-trained models from the computer science were applied for the 1 time to support joint space-time
st
domain to accomplish inter-domain knowledge transfer learning from recorded video dataset of complex process
as identified in another publication in detail. More phenomena (e.g., melting, solidification, and laser-material
14
recently, these approaches have begun using AM datasets interactions) to detect part quality and classify operating
and models to conduct intra-domain knowledge transfer modes. To overcome data scarcity and synthetically
due to the growing availability of AM datasets and applied generate realistic AM design, process, and structure data
ML solutions. As the applied ML models become more for the learning tasks, generative adversarial networks
66
complex and advanced, their black-box nature leads to a (GANs) were applied in 2021. 78-80 Since 2023, self-
lack of transparency by limiting the explanations for model attention-driven transformer-based networks and their
predictions. As a result, explainable AI techniques have variants tailored to specific modalities (vision, 81,82 3D,
83
seen growing applications recently. These explainable videos, language, 85,86 vision-language ) have started to
31
87
84
learning techniques reveal the underlying patterns learned find applications in AM. The first applications of large
by the model that drive its decisions and predictions, vision and language models (VLMs) in AM marked a
thereby enhancing trust and adoption through increased significant increase in the parameter scale (millions in
transparency. This is usually accomplished through CNNs, billions in large language models [LLMs], hundreds
feature importance scoring, parameter visualizations, and of billions in generative pre-trained transformers [GPTs])
contribution analysis, thereby linking inner elements of of applied ML models. While these architectures bring
model architecture with its predictions and AM inputs. 69-72 opportunities to model complex phenomena and handle
Figure 6 highlights the emerging learning trends for AM multiple modalities, they also introduce challenges to
applications with a focus on recent years. meet data requirements and ensure model robustness.
The applications of ML in AM indicate a clear trend Figure 7 identifies the first application of prominent parent
toward complex and large-scale models in terms of their categories of ML algorithms in AM.

Volume 1 Issue 1 (2025) 7 doi: 10.36922/ESAM025040004