Machine learning-based additive manufacturing process optimization and quality prediction with photomicrography and tomography

Abstract

Process optimization and quality prediction are essential for promoting manufacturing quality and improving product reliability in Additive manufacturing (AM). However, the complex multi-physics processes of AM, involving process-material interactions, melt pool dynamics, phase transformation, and microstructure development, result in challenges for finding the optimal process conditions (especially of AM fabrications with advancing novel materials design) and forecasting the mechanical properties of the AM parts. With the ever-expanding data availability of AM by various sources (e.g., photomicrography and tomography), data-driven machine learning (ML) modeling has gained increasing attention in AM research. Compared to trial-and-error or computational approaches, ML modeling can bypass the complex physics-based model of AM, explore the feasible space efficiently, and reduce time and cost for process optimization and quality prediction by using computationally inexpensive algorithms. In this dissertation, two different AM technologies, namely, fused filament fabrication (FFF) and laser beam-powder bed fusion (L-PBF) processes, are investigated for adoptions of ML-based process optimization and quality prediction, utilizing the data from photomicrography and tomography, such as the optical microscope, scanning electron microscope (SEM), and X-ray computed tomography (XCT). The overarching goal of this dissertation is to develop innovative methodologies tailored for different objectives regarding efficient process optimization and accurate quality prediction for FFF and L-PBF processes based on photomicrography and tomography data. Specifically, three new methodologies are developed as follows: First, a data-driven nonparametric Bayesian framework is developed to unify nanocomposite design and part fabrication by integrating statistical learning and optimization into quality prediction and process optimization. It regulates the design and fabrication of specimens with different materials composition using the FFF process and ensures successful printings and accurate quality prediction while maintaining its design flexibility and customization. It can be easily extended to other materials and will have a profound influence on advancing their applications and innovations in various industries. Second, an ML-driven framework is proposed to conduct an efficient and accurate defect classification for the L-PBF process by the distinct morphology and size of each defect type obtained from the XCT scans. It leverages the efficiency of low-resolution (LR) XCT scanning and ML to achieve improved defect inspection efficiency and classification accuracy simultaneously. The classification results from the proposed framework can be used as an effective method to identify the most detrimental defect type in the L-PBF parts to the fatigue performance and rectify the fabrication conditions of the L-PBF process for future manufacturing. Third, a cross-dimensional (3D/2D) defect matching process is proposed to identify the matches to the critical defects (i.e., defects at fracture origin) among numerous volumetric defects in the XCT scans of the L-PBF parts. It exploits the defect features extracted from SEM-based fractography to characterize the positions, sizes, and morphologies of the critical defect for identifying the volumetric defect in the XCT scans with the highest similarities to the critical defect as the match. Identifying the match to the critical defect not only reveals the important features of the critical defect, which contributes to the fatigue life of L-PBF parts but also opens the possibility of critical defect identification and fatigue life prediction non-destructively.