Page 51 - IJAMD-1-3

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

is crucial for enhancing model robustness, especially for
models that struggle with imbalanced datasets.
In terms of model performance evaluation, future
works will also include a detailed analysis of the impact
of dataset size and class distribution on model accuracy
and RMSE. Given the potentially significant influence of
imbalanced data on performance, this analysis will help
better understand the scalability and generalizability of the
models under different conditions. Data augmentation,
which may include techniques such as generative adversarial
networks or data synthesis algorithms to generate realistic
data, can be explored as a viable alternative to conducting
additional experiments. This approach is particularly
Figure 12. Pyrometer data from a layer of an additive-manufactured beneficial given the expensive and time-consuming nature
block with natural pores, and a single raster scan. of processing NiTi via AM, offering a cost-effective way
to both balance the data and increase dataset size without
at 81.92 bit/mm across a 400 × 400 mm area. As shown
in Figure 12, each layer’s pyrometer data is divided into the need for extensive physical experiments. Additionally,
raster scans, and the data around visible pores is truncated emerging state-of-the-art algorithms such as DrCIF and
to reduce noise and emphasize the underlying patterns. MINIROCKET, which can directly work with raw time‑
Raster scans passing through visible pores are labeled as series data, will be explored for their potential in porosity
porous, while others are classified as non-porous. The prediction. These models could offer improvements in
connection between temperature fluctuations and porosity both accuracy and speed by reducing the need for feature
is well established in the literature, 57,58 supporting the engineering. We aim to benchmark these algorithms
use of pyrometer data for this predictive modeling. The against traditional models to quantify improvements in
prediction of porosity in AM is strongly linked to the RMSE, with a target of reducing RMSE by at least 10% for
underlying physical mechanisms of melt pool behavior, high-porosity cases.
as captured through in situ pyrometer data. Temperature Moreover, while directly comparing the effects of
fluctuations within the melt pool play a crucial role in hyperparameters across ten diverse models is challenging
porosity formation, with rapid heating and cooling cycles due to the variation in hyperparameters (e.g., SVM uses
often resulting in the entrapment of gas or incomplete parameters such as C, gamma, and kernel type, while
fusion of metal powders. These temperature changes can RF requires tuning of the number of trees, minimum
cause melt pool instability, leading to defects like pores samples per split, and splitting criterion), a future study
due to insufficient material flow and uneven solidification. will focus on recommending models based on the ease of
Studies have shown that higher porosity levels tend to hyperparameter tuning. Some models, such as LR, require
occur when melt pool temperatures are inconsistent, as no hyperparameter tuning, while others, such as neural
rapid solidification can trap gas within the material. 57,68] By networks, involve complex hyperparameter settings. Such
monitoring these temperature fluctuations in real-time, ML study will involve identifying the key hyperparameters
models such as XGBoost can predict porosity and enable for each model (e.g., focusing on 3 – 4 most significant
proactive adjustments to process parameters, mitigating hyperparameters) and using techniques such as factor
defects and improving part quality. This approach allows analysis and principal component analysis to evaluate
for effective control of porosity, addressing a critical issue how many hyperparameters must be adjusted from default
in the mechanical properties and structural integrity of settings to achieve a specific performance threshold (e.g.,
AM parts. 90% accuracy). We will also factor in model training
time to provide a comprehensive comparison. This will
4. Future works enable more informed decision-making when selecting

To address the gaps identified in this study and further models based on both performance and ease of tuning,
improve the predictive performance of ML models in AM, filling a current gap in the literature. Herein, automated
several targeted directions are proposed. First, increasing feature selection was employed, in the future, an analysis
the dataset size is a priority. We plan to expand the dataset of how certain physical features directly contribute to
by at least 100%, focusing particularly on rare occurrences porosity prediction can be conducted, complementing the
such as high porosity. Collecting more data in these cases automated approach provided by TS-Fresh.

Volume 1 Issue 3 (2024) 45 doi: 10.36922/ijamd.4812