Automated Metrology and Machine Learning Predict Additive Manufacturing Quality

Applications of additive manufacturing (AM) are rapidly increasing because of advancements in AM processes, materials, and digital design tools. However, the inspection, measurement, and qualification of AM parts remains a key challenge. Most production AM parts undergo manual inspection, which is time consuming and expensive, limiting the economic viability of AM for many applications. This presentation covers two strategies for accelerating and improving part inspection and qualification processes.

First, we present methods for automatically measuring complex AM part geometries. The methods leverage computer vision and computational geometry to automatically identify critical features from part image data and measure the size, shape, and location of the identified features. This approach scales efficiently with part complexity and quantity, both of which tend to be high in AM.

Second, we present techniques for structuring manufacturing data for machine learning (ML). Design, metrology, and manufacturing process data, both categorical and continuous, are fused together to discover insights into AM processes, predict part dimensions, and classify parts as acceptable or unacceptable. We leverage semi-autonomous grid search hyperparameter tuning and cross-validation to explore different architectures and identify how much training data is required for making accurate predictions in different scenarios.

To demonstrate these methods, we manufactured 405 parts having three designs using three polymer materials and two identical machines. We automatically measured 2025 features and used this data, along with design and manufacturing information, to train a support vector regression model that accurately predicts as-manufactured part dimensions. The regressor is subsequently used to classify parts by applying two additional hyperparameters, achieving higher accuracy than direct classification models with greater explainability.

We find that the rich measurement data gathered using software automation, combined with ML prediction models, present opportunities to better understand geometry variability in production AM and lead to improved quality inspection systems.