Page 73 - ESAM-1-1

Engineering Science in
Additive Manufacturing ML in additive manufacturing

Figure 3. Interactions between data challenges in artificial intelligence-driven additive manufacturing

datasets, this trend was reversed in 2020 (Figure 4) and since
then has seen an increase in the applications of DL-based
models. By 2024, almost 70% of the ML applications in
AM were based on deep architectures capable of modeling
complex non-linear input-output relationships (e.g., fault
detection, anomaly classification, segmentation, and
synthetic data generation). Nonetheless, shallow learning
techniques remain useful for new applications (e.g., process
concerns, parameter, and material variations) with limited
datasets and elementary complexity.
The learning strategies in AM have been dominated
by supervised approaches where models are trained with
labeled data and evaluated against ground truth values
during the validation process. As presented in Figure 5, Figure 4. Shallow and DL in AM applications, 2015 – 2025. The plot
the initial years saw relative growth in applications of data was collected from 1250 research articles between January 2015
unsupervised learning (~25% in 2015 – ~40% in 2017) and January 2025. Several articles employed both shallow and DL-based
but the trend shrank afterward. This reflects the increasing techniques in their methodologies. The trend in 2025 was limited to
provision of labeled and annotated data in AM. Overall, research articles collected in January 2025.
the supervised learning approaches are the most popular Abbreviations: AM: Additive manufacturing; DL: Deep learning.
with over 80% of the applications falling in this category
in 2024. The remaining 20% of applications cover a realistic modeling of AM phenomenon (e.g., material
combination of unsupervised learning, self-supervised deposition, microstructure evolution) to support closed-
65
learning, semi-supervised learning, and reinforcement loop agent-based learning is challenging to accomplish.
learning (RL). Autoencoders and clustering algorithms The emerging trends among the learning strategies
that do not require annotated AM data represent major represent the efforts of the AM community to pivot towards
techniques among unsupervised learning strategies. The more robust models to enhance their applicability in the
applications of RL, where an agent learns to maximize the industry. The identified trends include physics-informed
reward in an environment by learning a sequence of steps, learning, knowledge transfer, explainable learning,
have been growing in AM since 2021. Only a handful ensemble-based learning, and active learning. Physics-
64
of works have utilized RL for process-based concerns as informed learning groups context-aware and engineering-

Volume 1 Issue 1 (2025) 6 doi: 10.36922/ESAM025040004