A Machine Learning Approach to Robotic Additive Manufacturing of UV-Curable Polymers Using Direct Ink Writing

This thesis presents the design and implementation of a robotic additive manufacturing system that uses ultraviolet (UV)-curable thermoset polymers. Its design considers future applications involving free-standing 3D printing by means of partial UV curing and the fabrication of samples that are reinforced with fillers or fibers to manufacture complex-shape objects. The proposed setup integrates a custom-built extruder with a UR5e collaborative manipulator. The capabilities of the system were demonstrated using Anycubic resin formulations containing fumed silica (FS) at varying weight fractions from 2.8 to 8 wt%. To fully cure the specimens after fabrication, a UV chamber was used. Then, measurements of the outside diameter (OD), height, and wall width of the cylindrical samples were taken with a digital caliper. The second contribution of this thesis is the design of a predictive approach to estimate the accuracy of the samples and remove expensive trial and error from the printing process using this proposed system. The printing parameters values were recorded and used as inputs to train a feedforward neural network (FNN) to predict the OD, height, and wall width of fully cured samples. This model was tested using 20 % of the total amount of data points. The FNN proved to be an accurate approach with the potential to save time and money when estimating the quality of manufactured objects if the printing parameter data and materials characterization data are available. As an alternative to the manual measurements, the printed samples were 3D scanned to investigate their quality. Then, iterative closest point algorithm (ICP) allowed to compare the ideal computer aided design (CAD) model with the manufactured sample. Using the root mean square error (RMSE) value from this comparison, an identical FNN architecture was trained to predict printed sample RMSE instead. It was found that the average RMSE value after testing increased to 0.7918 mm from 0.38 mm, decreasing the accuracy of the model. This approach is desired when the samples have more complex shapes. Nevertheless, further investigations remain to be done in this area to achieve high accuracy results and this remains the focus of our future work