Page 30 - ESAM-1-4

Engineering Science in
Additive Manufacturing Machine learning for biomedical metal AM

environments. This chapter elaborates on the application The universality of this data-driven approach was further
39
of ML in the forward prediction of these attributes for validated across diverse material systems. Gor et al.
68
AM-fabricated biomedical metals, demonstrating how conducted a systematic comparison of multiple models for
data-driven modeling supports performance forecasting. predicting the density of 316L stainless steel, with artificial
neural networks (ANN) and SVM demonstrating superior
2.2. Macrostructure quality forward prediction performance, achieving R² values of 0.95 and 0.92,
In additively manufactured biomedical metal, the respectively, for density prediction (Figure 4A and B).
macrostructure quality of parts forms the foundation for Collectively, these studies demonstrate that diverse
their successful application. Among these, density and ML models can achieve high predictive accuracy for
surface roughness are the two most critical macrostructure densification behavior in LB-PBF. By establishing precise
quality indicators. ML models enable precise forward quantitative links between process parameters and relative
prediction of these indicators by learning correlations density, these data-driven approaches offer a reliable
between process parameters and macrostructural strategy to avoid high-porosity process conditions, thereby
properties. enabling the direct fabrication of highly dense parts.

2.2.1. Density 2.2.2. Surface roughness
Density serves as a core metric for assessing internal Surface roughness is a critical factor determining
defects in additively manufactured parts. High density is product quality. In the engineering field, excessive
essential for ensuring the superior mechanical properties surface roughness not only directly leads to a shortened
and long-term service safety of biomedical metals. Early product lifespan but also significantly impairs
64
studies such as Bartolomeu et al. employed traditional mechanical properties such as tensile strength and
65
statistical models like response surface methodology. These fatigue strength. This degradation of macrostructure
successfully established quantitative relationships between properties is closely related to adverse effects on the
LB-PBF process parameters and density for Ti-6Al-4V, material’s microstructure. 69,70 For biomedical metals,
revealing the significance of parameter interactions and surface roughness directly regulates biocompatibility,
demonstrating preliminary validation of the feasibility of influencing cell adhesion, proliferation, differentiation
establishing quantitative process-density mapping through (osseointegration), and antibacterial performance. 71,72
data-driven methods. However, excessively high roughness can also become an
initiation site for fatigue cracks, posing a threat to long-
As datasets expand and algorithms advance, more
sophisticated ML models demonstrate superior predictive term service safety. Given the complex and critical multi-
performance. ML models, utilizing process parameters as dimensional impact of surface roughness on product
performance, its stringent control is essential.
inputs, can accurately predict the relative density of final
parts, thereby enabling the avoidance of high-porosity Consequently, various ML models have been developed
process windows before manufacturing. For instance, to predict surface roughness for different AM processes,
Maitra et al. constructed comprehensive Ti-6Al-4V aiming to achieve proactive control. Koo et al. employed
73
66
LB-PBF data set from 48 publications, employing GPR random forest regression, extreme gradient boosting
models to predict densification behavior. The GPR (XGBoost), and SVR models to predict the down-skin
model achieved a remarkably low MAE of 1.12%, and surface roughness in LB-PBF. The input features included
its outstanding engineering applicability was validated laser power, scanning speed, layer thickness, hatching
through actual printing tests (MAE = 0.27%). In another distance, and overhang angle. Among these models, the
study by Jiang et al., a dataset comprising 63 parameter XGBoost model demonstrated the best performance
67
trials was constructed with the aim of improving the (R = 0.63). For wire arc AM (WAAM), which typically
2
density and mechanical properties of the high-entropy involves larger thermal input and a more unstable process,
alloy Ti₁.₅Nb₁Ta₀.₅Zr₁Mo₀.₅ (TNTZM), a preferred clinical different approaches have been explored. Xia et al.
74
alloy, during the LB-PBF process. Using laser power and utilized a genetic algorithm-optimized adaptive neuro-
scanning speed as inputs, the AdaBoost model achieved fuzzy inference system (GA-ANFIS) model to predict
the best performance on the test set (R² = 0.85, RMSE surface roughness. This GA-ANFIS model achieved a
2
= 0.37). This approach successfully produced TNTZM very high prediction accuracy (R = 0.94), which was
samples with a density of 99.9%. Subsequent heat verified to be highly consistent with actual measurements
treatment increased the yield strength by over 150 MPa through a laser vision scanning system. This indicates that
while maintaining approximately 50% ductility, validating the GA-ANFIS model, combining the interpretability of
the predictive accuracy of ML model. fuzzy logic and powerful non-linear fitting capability, can

Volume 1 Issue 4 (2025) 8 doi: 10.36922/ESAM025440031