Page 52 - IJAMD-1-3

International Journal of AI for
Materials and Design
Metal AM porosity prediction using ML

By following these quantitative goals and exploring Acknowledgements
new algorithms, we expect to significantly improve the
performance of the ML models used for porosity prediction None.
in AM. These directions not only address the reviewer’s Funding
concerns about the vagueness of the original future work
section but also provide clear, measurable objectives aimed This publication reflects research supported by a research
at improving both accuracy and RMSE while considering grant from Science Foundation Ireland (SFI) under grant
practical aspects such as hyperparameter tuning and number 16/RC/3872, 21/RC/10295_P2, and co-funded by
computational efficiency. the European Regional Development Fund.
5. Conclusion Conflict of interest
This study addresses a knowledge gap by developing and Dermot Brabazon is an Editorial Board Member of this
implementing an in situ method for porosity prediction journal and Guest Editor of this special issue, but was not in
in additively manufactured parts. We approached the any way involved in the editorial and peer-review process
problem in two steps: first, classifying each layer as either conducted for this paper, directly or indirectly. Separately,
low or high porosity, and second, predicting the exact other authors declared that they have no known competing
porosity level using two separate regression models. financial interests or personal relationships that could have
Herein, the ability of 10 ML models to predict the porosity influenced the work reported in this paper.
of NiTi additively manufactured parts was compared
in terms of RMSE, accuracy, and speed. The dataset was Author contributions
collected from the Aconite MINI AM machine which Conceptualization: Vivek Mahato, Annalina Caputo,
consisted of pyrometer data. The porosity of additively Dermot Brabazon
manufactured parts was predicted using our experimental Formal analysis: Vivek Mahato
findings, highlighting several important conclusions and Investigation: Vivek Mahato, Annalina Caputo
suggesting future directions for research. We found that Methodology: Vivek Mahato, Annalina Caputo, Dermot
pyrometer data effectively captures relevant information Brabazon
about the underlying AM process, which can be valuable Writing–original draft: Vivek Mahato
for analysis, though it requires careful preprocessing to Writing–review & editing: Suman Chatterjee, Anesu
reduce noise and clean the data. By using an appropriate Nyabadza, Annalina Caputo, Dermot Brabazon
ML model, such as an XT regressor, we achieved strong
predictive performance for these tasks. Specifically, Ethics approval and consent to participate
classifying a layer as low (<1%) or high (≥1%) porosity Not applicable.
proved to be a relatively straightforward classification task,
whereas predicting the exact porosity percentage was more Consent for publication
challenging.
Not applicable.
In our experiments, the XT regressor achieved an RMSE
of 0.0367 on the “low” porosity dataset (479 observations) Availability of data
but a slightly higher RMSE of 0.108 on the “high” porosity
dataset (107 observations), which had fewer data points. Data is available from the corresponding author upon
When we undersampled the “low” dataset to match the size reasonable request.
of the “high” dataset, the model’s performance dropped, References
resulting in an RMSE of 0.165. These findings underscore
the importance of having large and balanced datasets, 1. Mahato V, Obeidi MA, Brabazon D, Cunningham P.
especially for rare occurrences such as high porosity, to Detecting voids in 3D printing using melt pool time series
achieve strong performance in regression tasks. data. J Intell Manuf. 2022;33:842-852.
doi: 10.1007/s10845-020-01694-8
For additional assessment of the ML algorithms’
performance in AM, more data on rare occurrences (i.e., 2. Herzog T, Brandt M, Trinchi A, Sola A, Molotnikov A.
high porosity) can be explored in the future. Further Process monitoring and machine learning for defect
analyses on the effects of the dataset size and class detection in laser-based metal additive manufacturing.
distribution on the performance of the ML models should J Intell Manuf. 2024;35(4):1407-1437.
also be examined. doi: 10.1007/s10845-023-02119-y

Volume 1 Issue 3 (2024) 46 doi: 10.36922/ijamd.4812