J. Killgore
National Institute of Standards and Technology,
United States
Keywords: vat photopolymerization, 3D printing, digital light processing, machine learning, convolutional neural network, generative adversarial network
Summary:
Digital light processing is a form of vat photopolymerization additive manufacturing where a digitally projected image locally photocures a liquid resin into a programmed solid. The method has applications ranging from military to automotive to medical, all increasingly relying on accurate, rapid production of high resolution, high performance parts. In digital light processing, voxel scale interactions are the critical disconnect between the resolution of the light engine (e.g. pixel size/pitch) and the resolution of a resultant 3D printed part (e.g. voxel accuracy). Over and under polymerization relative to the input photomask can occur due to numerous phenomena related to the non-ideality of the light source and heat and mass transfer in the reacting resin. While the process could be modeled directly from first principles, the parameter space becomes quite large and the input parameters are conversion dependent and difficult or impossible to directly measure. Here, we adopt a machine learning approach to train neural network models to predict resultant voxel patterns given arbitrary input photomasks. The training data are derived from laser scanning confocal microscopy height maps of voxels printed from randomly gray-scaled 8x8 pixel submasks. The gray scaling of individual pixels varies the intensity of local illumination, and thus the local reaction rate. The full training data consist of >50,000 pixel interactions, all imaged at micron-scale precision. We compare model prediction based on neural networks (NNs) with increasing level of sophistication, including a fully connected linear layer, a deep convolutional NN, and a conditional generative adversarial network (cGAN). The models are able to provide remarkably accurate predictions of complex voxel scale patterns. The models predict both lateral and thickness dimensions of the printed layers. With the model in hand, over and under polymerization of any arbitrary mask can be predicted and optimal mask design can be engineered.