Active Machine Learning for Projection Multi-Photon Lithography

<p dir="ltr">Additive manufacturing (AM), or 3D printing, has reshaped the manufacturing industry. By adding rather than subtracting material, 3D printing offers greater freedom of design, more rapid prototyping, lower waste, and many more benefits compared to traditional manufacturing methods. AM methods have been developed for applications of all sizes. At the smallest scales, multi-photon polymerization is the predominant method. Multi-photon polymerization is traditionally a direct laser writing process which allows for the 3D printing of polymer structures with submicron features. The nanoscale resolution and vast design space of multi-photon polymerization have led to its application in a broad range of research-related fields. Yet, its slower speeds and limited scalability have hindered its expansion into industry applications. As a result, recent research has prioritized increasing the scalability of multi-photon polymerization. A rapid projection multi-photon 3D printing process has recently been developed which parallelizes the printing process by projecting images to print entire layers of 3D structures at once. Advancement of this new projection multi-photon lithography (PMPL) method requires further understanding of the photopolymerization reaction and its behavior during the printing process. This can be enabled through both mechanistic models and machine learning models. This work investigates these avenues for improvement of PMPL. The structure of this dissertation is as follows. Chapter 1 provides the fundamentals and background information for traditional multi-photon lithography, PMPL, and machine learning in AM. Chapter 2 discusses finite difference modeling of the photopolymerization reaction. A model is introduced to describe the photochemical processes in two-photon lithography, including both excited-state kinetics and polymerization kinetics. The model is then used to analyze the effects of different process parameters and chemical parameters on the 3D printing process. Chapter 3 presents an active machine learning framework for optimization of 2D geometric accuracy in PMPL. The framework leverages a Gaussian process regression machine learning model within a Bayesian optimization scheme to guide efficient experimentation. The framework is tested on several simple 2D shapes and is shown to reduce geometric errors to within measurement accuracy in just four iterations. Chapter 4 discusses the use of this Bayesian optimization framework to optimize grayscale patterns for improvement of layer-height accuracy in PMPL. A simplified mechanistic model of the projection and polymerization process is incorporated for generation of grayscale patterns, and an in-situ quantitative phase imaging method is used for data collection to allow for fully autonomous optimization. Tests on representative 2D layers demonstrate that the method achieves uniform, accurate layers in under 300 prints. Chapter 5 summarizes the dissertation and discusses future work, particularly, improvement of the grayscale patterning methods and the extension of machine learning for optimization of fully 3D structures.</p>