A machine learning approach to establish a novel 3D printing method with PLA

Dan Luo, Dechen Chen, Weiguo Xu, and Guanqi Zhu
Department of Architecture, Tsinghua University, Beijing, China

Dan Luo, Weiguo Xu, Dechen Chen and Guanqi Zhu. A machine learning approach to establish a novel 3D printing method with PLA[C]. Proceedings of the 22nd International Conference on Advances in Materials & Processing Technologies (AMPT 2019)

Abstract. PLA (polylactide) is the most established 3D printing material that has been wildly adopted in commercial desktop additive manufacturing platform nowadays. Though the mechanism of each PLA printing system may be different, the material properties those system utilize for printing with PLA is fundamentally the same: that is the rapid thermoplastic transition from amorphous state to solid once been extruded. The form of printed outcome is the direct result of the extruding path. However, while the path of the extruder can be clearly planned, owning to the week intermolecular forces, after extruded the deformation of the material before reaching fully solid state is a dedicated balance between the movement of the extruder, material properties, temperature that has no stablished model to describe.  Thus, the deformation of the material owning to the amorphous property of PLA after been extruded is always an obstacle to the accuracy of printing that needs to be minimized in the past.

However, with the recent progress in machining learning, given enough training samples, it is possible to build a model without explicit knowledge of the mechanism behind the material performance. Thus, this paper presents an exploratory research that established a neural networks model that directly maps the GCode that controls a standard desktop PLA printer to the target printed curve that spans over two end support and draped into target shape, utilizing the amorphous property of PLA that has no established material model to describe. With such model, we are able to develop a 3D printing method that are capable of adapting a standard FDM 3D printer for printing cantilevering spatial wireframe, utilizing the deformation of PLA during the transition from amorphous to solid phase after extruded.

The workflow of the research started with building an automatic testing system that can automatically conduct material test in large quantity continuously. The results of the experiments are captured on image. Then, an image processing system is developed to format the images into features describing the target curve and matched with corresponding GCode, and formatted into datasets. The dataset is used for training of two neural network model that describe two-way mapping between the GCode controlling the printer and the printed spatial curve.