Page 82 - ESAM-1-1

Engineering Science in
Additive Manufacturing ML in additive manufacturing

which are often challenging to implement in practical asset (i.e., data and model), thus requiring expertise from
production settings. Therefore, this paper proposes and both AM and AI domains. A recent study indicated that
formulates industrial deployability, underpinning the crucial reproducibility in this domain is unsatisfactory due to
characteristics of AI-driven AM systems for real-world missing information in the literature, especially for data and
deployments. Industrial deployability encompasses seven model information. Therefore, researchers must develop
147
key elements: cost-efficiency, reproducibility, reusability, frameworks that organize AI-driven AM knowledge and
data privacy and safety, compliance with standards, guide reproduction. For example, Xie et al. proposed a
147
explainability, and human-AI teaming (Figure 11). reproducibility investigation pipeline and reproducibility
checklist to assist readers in extracting reproducibility
7.1. Cost-efficiency details from publications and navigating the replication
Cost-efficiency promotes industrial adoption by reducing process.
the infrastructure and computational costs required
to construct and operate the system. On the one hand, 7.3. Reusability
AI-driven AM systems aim at increasing part quality Reusability concerns whether the data and knowledge
and reducing waste to gain more profit in production. from established AI-driven AM systems can be adapted
On the other hand, advanced ML architectures such as and reused for a new task or newly developed AI-driven
transformers and multi-modal learning substantially AM systems. The performance of ML models is usually
increase the infrastructure and computational costs for compromised when domain shift occurs, which refers
deployment. Cost-efficiency must be enhanced by removing to any changes in the AM assets, including the machine,
redundancy in sensor configuration, data acquisition, and materials, process parameters, and sensors. In real
model architecture to improve profitability. For example, production environments, it is highly costly for industrial
the number of sensors in an in situ monitoring system practitioners to collect data and train new models from
can be reduced without performance compromise using scratch for every domain shift occurring on the shop
optimization and knowledge transfer. 144,145 floor or for new machines. However, knowledge from
AI-driven AM systems can be reused due to similar
7.2. Reproducibility physical phenomena and system setups. For example,
AI-driven AM systems proven effective in laboratories knowledge transfer can be conducted from established
must be replicable for industries to deploy on their shop systems to newly developed systems to avoid expensive
floor. Reproducibility, a major challenge for scientific data collection. 89,148 Reusability is distinguished from
discovery and industrial deployment in various domains, reproducibility because it addresses the scalability for new
describes whether comparable performance can be tasks (e.g., printing new geometry or material), whereas
obtained by a different group with a different experiment reproducibility concerns the necessary information to
setup. 146,147 Achieving high reproducibility in AI-driven replicate the setup of an established system. Technically,
AM is difficult because it concerns both the AM asset reusability covers knowledge transfer methods and
(i.e., AM machine and monitoring system) and the AI reproducibility covers information modeling methods.

Figure 11. Seven key elements of industrial deployability for AI-driven AM
Abbreviations: AI: Artificial intelligence; AM: Additive manufacturing.

Volume 1 Issue 1 (2025) 15 doi: 10.36922/ESAM025040004