Leveraging Machine Learning for Porosity Prediction in AM using FDM for Pretrained Models and Process Development

Due to the numerous independent parameters involved in additive manufacturing, 3D printing often delivers different quality of prints and requires an expensive trial-and-error approach before finding the optimal combination of the numerous and independent input variables. Machine learning is, therefore, an ideal solution to this nonlinear problem, and provides an informed guess on printing parameters based on a minimal set of experiments. Using the case example of Fused Deposition Modeling, and, as a proof of concept, examining the porosity defect, a machine learning powered process is developed to predict the porosity defect's occurrence. It also facilitates the determination of the types of combinations of printing variables to ensure minimal defects. Specimens were 3D printed and CT-scanned. Raw datasets were collected in the form of grayscale image files (around 7,300 images) from the CT-scan. A machine learning image classifier was developed and trained to sort exploitable images from defective ones. To preprocess information for the classifier, intelligent scripts were created to extract porosity features. A Multi-Layer Perceptron (MLP) was then developed and trained to predict porosity across the specimens’ height.. Given the size of the dataset and input features, the model's accuracy has proved to be optimal: The perceptron was able to predict reasonable porosity values for established and unknown combinations of input variables for two different sets of specimens. Finally, a scalability study was conducted to establish the impact of scaling on defect formation and the prediction of 3D printed parts.