Machine learning for continuous liquid interface production: Printing speed modelling

Abstract

Continuous Liquid Interface Production (CLIP), a variant of vat photopolymerization additive manufacturing, can achieve build speeds an order of magnitude faster than conventional layer-by-layer stereolithography process. However, identification of the proper continuous printing speed remains a grand challenge in the process planning. To successfully print a part continuously, the printing speed needs to be carefully adjusted and calibrated for the given geometry. In this paper, we investigate machine learning techniques for modeling and predicting the proper printing speed in the CLIP process. The synthetic dataset is generated by physics-based simulations. An experimental dataset is constructed for training the machine learning models to find the appropriate speed range and the optimum speed. Conventional machine learning techniques including Decision Tree, Naïve Bayes, K Nearest Neighbors, and Support Vector Machine (SVM), ensemble methods including Random Forest, Gradient Boosting, and Adaboosting, and the deep learning approach Siamese Network are tested and compared. Experimental results validate the effectiveness of these machine learning models and show that the Siamese Network model gives the highest accuracy.