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

Area of Science:

• Materials Science
• Computer Science
• Mechanical Engineering

Background:

• Additive manufacturing processes like Fused Deposition Modeling (FDM) have numerous parameters, leading to inconsistent quality and requiring extensive trial-and-error for optimization.
• Machine learning (ML) offers a data-driven approach to predict and optimize these complex, non-linear relationships.

Purpose of the Study:

• To develop an ML-powered pipeline for predicting porosity defects in FDM-printed parts.
• To assess the impact of geometrical scaling on defect prediction accuracy.
• To quantify process variability for enhanced quality control.

Main Methods:

• 3D printing of specimens at two geometrical scales, followed by CT-scanning to generate image datasets.
• Training an ML image classifier to identify defective vs. exploitable prints.
• Developing preprocessing scripts to extract porosity features.
• Training a multi-layer perceptron (MLP) model on extracted features.
• Implementing a grouped k-fold cross-validation protocol for robust model evaluation.

Main Results:

• The image classifier achieved over 97% accuracy in distinguishing defective from exploitable images.
• MLP model accuracy increased from 54.4% (small scale) to 77.6% (large scale) with larger datasets.
• Repeatability studies quantified intrinsic process variability, with an average porosity standard deviation of 0.47%.

Conclusions:

• An ML pipeline can effectively predict porosity defects in FDM parts, reducing the need for manual optimization.
• Geometrical scaling significantly impacts defect formation and the accuracy of ML-based prediction models.
• The developed methodology provides a framework for robust quality control in additive manufacturing.