Page 89 - MSAM-4-3

Materials Science in Additive Manufacturing Interpretable GP melt track prediction

3.3. Morphological prediction defect types, and may also affect other defect categories.

Following physical constraint-guided DGP model The issue could be due to the presence of composite defect
predictions of melt track dimensions, the deviation morphologies within the melt tracks, as exemplified by
feature (Equation XIX) was used to predict the melt track the three typical cases displayed in Figure 19. Specifically,
morphology (Figure 17). This approach demonstrates the composite necking-protrusion morphology can
robust predictive performance based on quantification by produce an average melt track width similar to that of the
accuracy (91.89%), precision, recall, and F1-score metrics. regularity category in the final sampling results. Since the
classifier relies primarily on global features and overlooks
Analysis of evaluation metrics (Figure 17A) and the local feature mutations, classification boundaries become
confusion matrix (Figure 17B) revealed sample imbalance blurred, leading to misses (leakage) and false positives.
across defect categories. To address this, two-stage SMOTE Consequently, enhancing the classifier’s ability to
oversampling was applied to the five defect classes, accurately recognize multiple co-occurring defects within
consequently generating improved predictions (Figure 18). a single melt track remains a significant challenge.
A comparison between Figures 17A and 18A reveals These findings further indicate that samples from
significant recall improvements of 5.3% and 11.6% in the different morphology categories occupy overlapping
hump and protrusion categories, respectively. In addition, regions in the feature space, leading to confusion and
the standard deviation decreased across all categories, while ambiguous classification boundaries.
the model’s predictive consistency improved. Although
the model exhibited reduced precision and fluctuations in To understand the importance of different deviations
accuracy, its substantial enhancement in defect detection for the class prediction of melt tracks, a feature importance
recall demonstrates satisfactory overall performance. 38 analysis method based on permutation importance was
Similarly, a comparison between Figures 17B and 18B used to verify whether the physical model of the theoretical
reveals a notably high leakage rate (exceeding 16%) in the deviations of melt tracks can effectively distinguish the
regularity category, which is prone to confusion with other classes of different melt track morphologies. Figure 20A
demonstrates the importance of different deviations in
the prediction of categories, and Figure 20B demonstrates
A B C the volumetric deviations on the upper surface of the
melt track for different categories. These findings indicate
distributional overlap among different morphology
categories in the feature space, resulting in ambiguous
classification boundaries and misclassifications. However,
volumetric deviation demonstrates the discriminative
capability for melt track morphology to some extent,
with its high feature importance confirming its role as
a critical classifier for morphology differentiation. The
Figure 19. Composite morphology. (A) necking-protrusion. (B) distortion- corresponding melt track deviation and height deviation
necking-protrusion. (C) hump-protrusion. Scale bar: 200 μm values become particularly important for distortion and

A B

Figure 20. Analysis of classification results: (A) importance of features, and (B) melt track volume deviation

Volume 4 Issue 3 (2025) 15 doi: 10.36922/MSAM025200030