Convolutional autoencoder frameworks for projection multi-photon 3D printing

Abstract

Projection multi-photon 3D printing is an emerging technique for fabricating micro-nano structures at exceptionally high speeds. It leverages the use of a digital micromirror (DMD) to project and print entire 2D layers at once, offering higher throughput and scalability than conventional point-by-point laser scanning. While two photon polymerization is widely regarded as an outstanding method for achieving high dimensional accuracy at the nanoscale, the projection aspect introduces a new set of challenges, such as under-printing due to oxygen inhibition. The inherently complex photopolymerization dynamics make it difficult to model and simulate efficiently. To address this, we introduce a data-driven methodology employing deep learning to build a surrogate model of the printing process and an inverse model for 2D DMD pattern optimization to achieve desirable printed shapes. By printing diverse shapes morphed by various parametrization schemes, we built a dataset for training convolutional encoder-decoder (autoencoder) neural networks. The trained surrogate accurately maps input DMD patterns to their final printed geometries, capturing nonlinearities introduced by process physics. Inverting the inputs and outputs further enabled us to train an inverse model for generating pre-compensated DMD patterns to print desirable target geometries. Experimental findings demonstrate that this deep learning approach accurately predicts printed outputs and enhances dimensional accuracy in the printing of 2D layers. Our results reveal a viable approach to overcome inhibition-induced constraints, enabling more accurate projection-based multi-photon printing at the micro and nanoscale.