Page 35 - ESAM-1-4

Engineering Science in
Additive Manufacturing Machine learning for biomedical metal AM

for predicting the fatigue behavior of actual components and avoiding defects before production is a critical task for
with geometrical features, highlighting its substantial ensuring product quality. In this context, ML, particularly
practical engineering value. Regarding direct prediction its important subfield, has demonstrated significant
based on process parameters, Zhang et al. constructed an potential in key tasks such as defect identification,
90
ANFIS model for LB-PBF 316L stainless steel. The model classification, and generation.
successfully predicted high-cycle fatigue life using process In the diagnosis and prediction of defect types, Lee
parameters and post-processing parameters in conjunction et al. established an explainable ML framework for the
94
with the maximum cyclic stress, maintaining prediction laser metal deposition process. The study used GPR to
errors within an acceptable range. Similarly, Shen et al. predict porosity and geometric dimensions, and an SVM
91
employed feature engineering to optimize a multilayer to classify defect types (e.g., gas pores, keyholes, and lack
perceptron model for Ti-6Al-4V alloy. By incorporating of fusion), achieving an overall accuracy exceeding 0.93.
defect characteristics obtained from micro-CT scans, Through SHAP analysis, the influence weight of various
they established a high-precision fatigue life prediction process parameters on defect formation was systematically
model and explicitly identified defect location as the quantified. Results indicated that the powder feeding
most critical factor influencing life. Evidently, the ML rate had the greatest impact on the deposited height,
prediction of fatigue life is evolving from a reliance solely while laser power was the most critical factor controlling
on process parameters toward an integrated framework 95
that incorporates mechanical loading conditions with key porosity. Similarly, Gui et al. designed and fabricated 32
defect characteristics such as size, shape, and location. sets of S30C samples with different process parameters
and conducted actual inspection and classification of
2.6. Defect forward prediction internal defects using X-ray CT and SEM. By employing
quantitative surface morphology parameters as defect
Common defects in AM, such as porosity, lack of fusion, detection criteria, an SVM model was utilized to predict
and cracks (Figure 9), are key risk sources that compromise defects based on process parameters (current, scanning
the reliability of metals 92,93 and can lead to clinical failure. speed, and line offset), achieving an accuracy of 0.95 and
These defects not only act as stress concentration sites, an F1 score of 0.73.
significantly reducing the fatigue strength and fracture
toughness of the material, but in biodegradable metals, Furthermore, Du et al. integrated purely data-driven
96
they may also accelerate localized corrosion, leading methods with physical laws to reveal the underlying
to premature mechanical degradation or the release of mechanisms of defect formation and establish high-
unintended metal ions. Therefore, proactively predicting precision prediction models. Addressing balling defects in

A B

C

Figure 9. Common defects in additive manufacturing. (A) Lack of powder fusion; (B) Crack; (C) Formation of keyhole. 93
92

Volume 1 Issue 4 (2025) 13 doi: 10.36922/ESAM025440031