Learning and Recalibration With Small Sets of Shapes for 3D Printing Poster.pdf

Calibration, or accuracy control, of cyber-physical additive manufacturing systems relies on predictive models for geometric shape deformation. However, learning predictive models is made difficult by the wide variety of possible process conditions and shapes. In addition, resource constraints limit the manufacture of test shapes, which further impedes learning of deformation models for new shape varieties. A methodology that can make full use of data collected on different shapes and reduce the haphazard aspect of traditional learning techniques is necessary in this context. We develop a Bayesian methodology that effectively combines deformation data and models from a small sample of disparate, previously manufactured shapes for the systematic construction of predictive deformation models of a broad class of new shapes. The power and simplicity of this general methodology is demonstrated with illustrative case studies on learning in-plane deformation models for the straight edges in a shape using only data and models for flat cylinders and a single regular pentagon. Ultimately, our Bayesian learning methodology facilitates deformation modeling in general, and enables recalibration of ineffective deformation models with little further experimentation. The broader impact is that it can help advance a smart cloud-based 3D printing app with the potential of immediate practical application.