Page 33 - ESAM-1-4

Engineering Science in
Additive Manufacturing Machine learning for biomedical metal AM

cross-printer knowledge transfer across three industrial of each input feature to model predictions, thereby
scenarios and significantly reducing the experimental revealing the influence mechanisms of process parameters
burden for process qualification. In addition, Fang et al., on strength and ductility. As shown in Figure 7, the
83
as illustrated in Figure 6, integrated a validated thermal MechProNet framework developed by Akbari et al.
86
model with a 1D convolutional neural network (CNN) integrated 1600 data points and employed models such
to analyze thermal history data, accurately predicting the as random forest to achieve high-precision predictions
mechanical properties, specifically the UTS, yield stress, (R > 0.90) for yield strength, tensile strength, and other
2
and failure stress of DED Inconel 718 thin walls and properties. Concurrently, SHAP analysis systematically
demonstrating high predictive accuracy for UTS with a test examined the influence of multiple AM factors on
set R of 0.67. The thermal history-performance prediction mechanical properties such as yield strength and UTS
2
framework established in this study is equally applicable (Figure 8). The analysis indicated that post-processing
to forecasting performance changes in biomedical metals conditions are the most significant factor affecting yield
during AM caused by complex thermal cycling. For the strength, followed by material thermophysical properties,
elastic modulus which is critical to biomedical implants, processing parameters, and the material itself.
Liu et al. developed a Young’s modulus prediction model SHAP analysis serves a dual purpose: it quantifies
84
for Ti-6Al-4V based on an ANFIS. This model achieved the importance of influencing factors and reveals the
a prediction error of only 0.66 GPa, far outperforming direction of their effects on mechanical properties. This
traditional theoretical models, and effectively quantified enhances the transparency of ML models and validates the
microstructure effects, providing a reliable foundation for consistency between predictions and underlying physical
precision control of implant modulus. mechanisms. Such interpretability not only improves
With the increasing application of ML models in model trustworthiness but also provides new insights
AM of biomedical metals, model interpretability has into understanding complex process–structure–property
become crucial for ensuring prediction reliability relationships in AM. By clarifying the influence mechanisms
and guiding process optimization. Shapley additive of key process parameters, it effectively guides the process
explanations (SHAP) analysis, one of the most widely used development of novel biomedical alloys, reduces trial-and-
interpretability methods, quantifies the contribution error costs, and accelerates clinical translation.
85

Figure 6. Prediction and characterization of additively manufactured metals by machine learning. Variational autoencoder training principle and latent
space feature maps for microstructure heterogeneity identification. 83
Abbreviations: AM: Additive manufacturing; IR: Infrared; PSP: Process-structure-properties; SEM: Scanning electron microscopy; UTS: Ultimate tensile
strength.

Volume 1 Issue 4 (2025) 11 doi: 10.36922/ESAM025440031