Page 28 - ESAM-1-4

Engineering Science in
Additive Manufacturing Machine learning for biomedical metal AM

similar characteristics, without the explicit definition and final properties defies description through simple
of the nature of these groups. Common algorithms models; (ii) inverse optimization of process parameters
include K-means, hierarchical clustering, generative inefficiencies: global optimization and multi-objective
adversarial networks, and self-organizing maps trade-off inefficiencies in high-dimensional parameter
(SOMs). Dimension reduction algorithms reduce data space; and (iii) quality control challenges: insufficient
dimensions to enhance visualization and processing real-time diagnostic and intervention capabilities
efficiency while preserving key features and lowering for identifying and addressing defects during the
complexity. These algorithms play a crucial role in manufacturing process. It is within this context that ML
handling high-dimensional datasets, particularly in emerges as a transformative tool, capable of uncovering
image processing tasks. hidden patterns from historical data, simulations, and
(iii) Semi-supervised learning: Semi-supervised learning real-time process monitoring. This data-driven approach
algorithms are positioned as an intermediate model enables the construction of precise predictive models and
between supervised and unsupervised learning efficient optimization strategies, thereby addressing the
approaches. This methodology employs the use of core bottlenecks in AM.
pseudo-labels or a limited amount of labeled data in Although existing reviews have improved our
conjunction with a substantial volume of unlabeled understanding of biomedical metals and ML applications
data for the purpose of performing data classification in AM, the research in this specific subfield is not
and achieving predictive outcomes. Key algorithms integrated. For instance, Bahl et al. systematically
53
include self-training classifier and graph convolutional reviewed the design, thermo-mechanical processing,
networks. This learning approach is particularly and performance evaluation of metastable β titanium
valuable when labeled data acquisition is costly but alloys for biomedical applications. Li et al. reviewed
54
unlabeled data are relatively accessible. the effects of various surface morphologies on the
(iv) Reinforcement learning: Reinforcement learning in vitro and in vivo performance of biomedical metallic
represents one of the more complex ML models. It materials, while delving into the underlying mechanisms
employs a trial-and-error approach to seek optimal linking surface topography to biological responses.
solutions within specific domains, learning through Wang et al. reviewed the research progress on Mg-based
55
positive reinforcement signals generated by its alloys as biomedical materials, with a particular
environment. Typical algorithms include Q-learning, focus on strategies for optimizing their mechanical
deep Q-network, policy gradient, and Markov properties and corrosion behavior. Existing reviews on
decision process. The field of reinforcement learning biomedical metallic materials seldom incorporate ML
has demonstrated significant potential in domains methodologies; they are more frequently mentioned
such as robot control and resource scheduling.
as future perspectives in the outlook section, such as
Model evaluation must be conducted on data Guo et al. specifically highlighted the prospects of ML
56
independent from the training set to avoid overfitting or in accelerating intelligent design, printing processes,
underfitting. Common evaluation methods include k-fold and performance prediction of powder-based three-
cross-validation. Evaluation metrics are selected based dimensional (3D) printed titanium alloys providing
on task type: regression problems commonly use root new insights for future research. For the field of
mean square error (RMSE), mean absolute error (MAE), ML-driven alloy design, Li et al. outlined the use of ML
57
and coefficient of determination (R ); classification in multi-property optimization of high-entropy alloys,
2
problems employ accuracy, recall, and F1 score; and while Hu et al. summarized ML’s role in establishing
58
clustering problems may utilize metrics like the Landis composition-processing-property linkages and enabling
index. Evaluation results can further optimize model inverse design. For ML in AM, Jin et al. elaborated on
59
hyperparameters to achieve optimal performance. 51,52 the specific applications of ML across different AM
processes and the construction of digital twins, and
1.3. ML-driven AM of biomedical metals Zhu et al. reviewed ML applications in condition
60
AM offers freedom for the personalization of biomedical monitoring and defect diagnosis for metal AM. Further
61
metals, while the highly nonlinear and dynamically expanding on the applications of ML in AM, Chen et al.
uncertain process poses significant challenges in reviewed the application of ML in AM from design and
achieving precise control over process-structure-property process optimization to in situ defect detection and
relationships. The challenges manifest across three post-process quality assessment and highlighted the
aspects: (i) forward prediction of properties difficulties: innovative integration of emerging ML techniques,
the intricate relationship between process parameters such as reinforcement learning and physics-informed

Volume 1 Issue 4 (2025) 6 doi: 10.36922/ESAM025440031