Page 42 - IJAMD-1-3

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

Table 1. (Continued) Next, we provide brief descriptions of our feature engineering
and summarizing tool and all the ML algorithms employed
No. Insights of research References
(using Python 3.11) in Section 3. In Section 4, we summarize
13 • The paper develops a predictive model for 31 our evaluation pipeline, the two ML tasks (classification and
porosity in additive manufacturing. regression) operated on the AM data, and the evaluation
• Focuses on high‑energy‑density regimes in
powder-bed fusion. metrics. The results are then presented in Section 5, with
• Discusses bubble formation and trapping conclusions and future work discussed at the end of the paper.
during solidification.
• Aims to control and reduce porosity in 2. Materials and methods
manufacturing processes.
• Highlights the significance of porosity on 2.1. Experimental setup
fatigue performance.
Rectangular objects were fabricated using the L-PBF process
Abbreviations: AM: Additive manufacturing; CT: Computed
tomography; DCNN: Deep convolutional neural network; in the present research work. The specimens were produced
DSPMs: Defect structure process maps; L-PBF: Laser powder-bed using the AconityMINI 3D laser powder-bed fusion-laser
fusion; ML: Machine learning. system (Aconity, Germany), which incorporates a 200-
Watt 1068-nanometer Ytterbium (Yb) fiber laser (Figure 2).
raw time-series data is essential, as it retains the information The laser system can attain up to 2000 mm/s velocities and
regarding the underlying process, which is sequential by possesses a minimum laser beam diameter of 32 mm. The
54
nature, that could be lost when using summary statistics “Skywriting” functionality was activated to fabricate the
(such as data mean and variance values). Classification or rectangular samples 10 × 10 × 10 mm (length, width, height),
3
regression based ML methods can be used to analyse time and each sample had identical dimensions to maintain the
series data. 39-46 The extent of data available for training uniformity of the study. The samples were produced on a NiTi
models can be increased via the combined efforts of substrate with a diameter of 125 mm. The L-PBF-fabricated
increased experimentation as well as ML data boosting samples may contain various anomalies in the manufactured
methods. 47-53 . parts, such as internal porosity, hot cracking, or the formation
55
In this research, the problem of detecting pores during of material aggregation on the component’s surface. The
the manufacturing process is examined via two different present study uses ML algorithms to identify and predict
tasks: predicting an AM block’s layer to be of low or high the internal porosity in the built-up samples. To oversee the
porosity (a binary classification task) and predicting the current manufacturing procedure, two in situ pyrometers
porosity percentage of each of the layers (a regression from Kleiber Infrared GmbH were used to record the melt
task). The capability to successfully predict such events pool temperature at 100 kHz at each x, y, z location.
can be translated into detecting faulty layers as they occur, The focus of this investigation is to predict the porosity
allowing the process to be stopped or corrected, resulting of an AM block’s layer based on its temperature time-series.
in significant waste reduction. The temperature time series Based on the porosity distribution manifested by the dataset,
is one of the main contributing factors to pore formation the plan is to target this problem as both classification and
which is captured via an in-line pyrometer. The resulting regression tasks. Classification predicts whether the layer
time-series data can capture the complex temporal is of low or high porosity (i.e., one binary classification).
interaction between the melt-pool temperature and the Subsequently, the regression models, one for low porosity
porosity of that zone. For this purpose, we have created a (0 – 1%) and another for high porosity (>1%) samples
dataset of naturally occurring pores (from a build of 10 mm predict the layer’s porosity percentage, a non-negative real
× 10 mm × 10 mm nickel-titanium [NiTi] metal blocks) for number. The threshold value (of 1%) was selected based
developing and analyzing the prediction and classification on a visible split in porosity distribution. The details of the
steps. The dataset reports the melting pool’s temperature aforementioned ML tasks and their evaluation metrics and
as the laser traces the part on the powder-bed bit-by-bit strategies are presented in the following subsections.
and layer-by-layer. Following a dataset analysis, we devised
a pipeline to balance the data using data augmentation. 2.2. Exploratory data analysis (NiTi×3B data)
Based on the dataset, we employed a package of strategies Employing AconityMINI, we constructed a batch of 10 mm
that take advantage of the sequential nature of these time- × 10 mm × 10 mm NiTi metal blocks with varying layer
series data to extract summary features that traditional ML thicknesses (30, 60, and 90 microns). The samples were
models can effectively optimize. fabricated according to a 2 design of experiments, resulting
3
The following section defines our system for capturing in eight sample fabrication permutations using the process
data and describes the used data augmentation technique. parameters of laser power (160 W and 200 W), scanning

Volume 1 Issue 3 (2024) 36 doi: 10.36922/ijamd.4812