Predicting Manufactured Shapes of a Projection Micro-Stereolithography Process via Convolutional Encoder-Decoder Networks

Projection micro-stereolithography (P-μSLA) processes have been widely utilized in three-dimensional (3D) digital fabrication. However, various uncertainties of a photopolymerization process often deteriorates the geometric accuracy of fabrication results. A predictive model that maps input shapes to actual outcomes in real-time would be immensely beneficial for designers and process engineers, permitting rapid design exploration through inexpensive trials-and-errors, such that optimal design parameters as well as optimal shape modification plan could be identified with only minimal waste of time, material, and labor. However, no computational model has ever succeeded in predicting such geometric inaccuracies to a reasonable precision. In this regard, we propose a novel idea of predicting output shapes from input projection patterns of a P-μSLA process via deep neural networks. To this end, a convolutional encoder-decoder network is proposed in this paper. The network takes a projection image as the input and returns a predicted shape after fabrication as the output. Cross-validation analyses showed the root-mean-square-error (RMSE) of 10.72 μm in average, indicating noticeable performance of the proposed convolutional encoder-decoder network.