Page 91 - MSAM-4-3

Materials Science in Additive Manufacturing Interpretable GP melt track prediction

Figure 22 displays the sensitivity of the melt track geometric features of the melt pool in real time. This study
characteristics to changes in the characteristics of the provides a reference for AM process monitoring and
melt pool. Changes in the convexity of the melt pool and defect prevention and control. Future research will further
the roundness of the melt track are strongly correlated investigate how single-track defects propagate through
with variations in melt track width, while melt track multilayer and multi-track interactions.
offset primarily responds to shifts in the melt pool’s
centroid position and area. In addition, melt track height Acknowledgments
fluctuations are dependent on alterations to the melt pool’s The authors would like to acknowledge financial support
aspect ratio and area. This observation validates the defect- from the National Natural Science Foundation of China
causation hypothesis proposed in Section 2.3. Analysis of under Grant Number 52175481 and the China Post-
the time-lag mutual information matrix further reveals doctoral Science Foundation under Grant Number
near-synchronized responses between melt track width 2023M743539.
and deviation changes, whereas height adjustments exhibit
a measurable temporal lag. Funding

4. Conclusion This work was supported in part by the National Natural
Science Foundation of China (grant no. 52175481) and in
The proposed DGP-p model establishes an interpretable part by the China Post-doctoral Science Foundation (grant
feature-mapping mechanism. Experimental results no. 2023M743539).
demonstrate that the physical kernel function enhances
intra-group sample variance within the same power Conflict of interest
category (from 0.456 to 0.547, an improvement of 19.9%) Xin Lin serves as the Editorial Board Member of the
while reducing inter-group covariance across different journal but was not in any way involved in the editorial and
power categories (from 0.256 to 0.176, a reduction of peer-review process conducted for this paper, directly or
31.3%), thereby validating its effectiveness in filtering non- indirectly. Other authors declare they have no competing
physical feature correlations. interests.
The model demonstrated superior performance under
single-sensor conditions, with average relative errors of Author contributions
9.89% (width), 10.21% (deviation), and 16.03% (height), Conceptualization: Xin Lin and Kunpeng Zhu
outperforming traditional machine learning methods. Formal analysis: Xin Lin
With a 9.8 ms inference time (GPU: 4070super) – Investigation: Xin Lin, Kunpeng Zhu, and Haodong Chen
slightly slower than the GP model’s 7.5 ms – it achieved Methodology: Shilin Liu, Xin Lin, Kunpeng Zhu, and
65.57% higher prediction accuracy. The melt track defect Jinrong Mao
classifier – combining volumetric deviation (weight = 0.39) Writing – original draft: Shilin Liu, Haodong Chen, and
and geometric features – attained 90.76% classification Xin Lin
accuracy. Writing – review and editing: Xin Lin
This study employed feature-sensitive analysis and
time-lag mutual information quantification. Notably, Ethics approval and consent to participate
there is a continuous effect of the dynamic behavior of the Not applicable.
melt pool on the geometrical characteristics of the melt
track: the melt track width is affected by the convexity Consent for publication
characteristics of the melt pool and the area. The deviation Not applicable.
of the melt track is affected by the position of the center of
mass and the area of the melt pool, while the height of the Availability of data
melt track is affected by the length-to-width ratio of the The raw/processed data required to reproduce these
melt pool and the area.
findings cannot be shared at the time of publication, as the
Overall, this study highlights that there is a continuous data also forms part of an ongoing study.
effect on the changes induced by these features, with
a significant lag effect on the changes induced by the References
melt pool aspect. Based on these findings, an effective 1. Khairallah SA, Martin AA, Lee JR, et al. Controlling
early warning mechanism can be established at the early interdependent meso-nanosecond dynamics and
stages of defect formation by monitoring changes in the defect generation in metal 3D printing. Science.

Volume 4 Issue 3 (2025) 17 doi: 10.36922/MSAM025200030