Metal 3D Printing Melt Pool Prediction Model
The major problems in laser-based metal 3D printing are porosities and irregular densities in the final printed piece. This leads to poor part quality and often causes parts to fail unpredictably. In this research, we developed numerous predictive models that take in the melt pool's temperature map and identify these errors in real time. This can then be used on the next layer to either remelt or exert more material on the next pass.
However, in this example, we will only focus on creating a neural network model to predict porosity formation. Although we prototyped a plethora of models, a neural network produced the best result. For training the network, a 9:1 cross-validation split for both the training and testing datasets was used to determine the effects of overfitting.
The collection on the left is an example of a melt pool after a threshold cutoff and normalization. The hottest temperature is at the center of a 40 x 40 input matrix.
Example of an original melt pool data
After training the network on normal melt pool data, we experimented with other augmented data sets to determine various relationships between the melt pool and porosity.
Binary matrix describing the outline of a melt pool
A 20 x 20 temperature map, only using a quarter from the 40 x 40
In the end, the neural network achieved an accuracy of 93.33 % on unseen test cases. The final architecture of the network is as follows:
40 x 40 input matrix with the hottest temperature in the center
A 3 x 3 sharpness filter is applied
Normalization of the 38 x 38 matrix as inputs
The first hidden layer of 75 nodes
The second hidden layer of 15 nodes
Binary output node: porosity or not porosity
In the future, we hope to implement this network in real-time and utilize it with an actual metal 3D printer