Content area
Laser-based metal additive manufacturing (AM) technologies, such as laser powder bed fusion (LBPF) and laser direct energy deposition (LDED), have been widely adopted in a wide range of industries including the aerospace, automotive, biomedical and energy sectors. Compared with conventional manufacturing processes, the AM processes involve a wide range of processing parameters including feedstock characteristics, laser power, spot size, scanning speed, layer thickness and hatch spacing, which are essential to the quality, properties, and performance of final products. The identification of processing windows from a vast process parameter space is a daunting task. Despite an increasing theoretical understanding and numerical simulations of the metal AM methods, the optimization of the AM processes is predominantly developed through a sequential and time-consuming trial-and-error approach, especially for new materials. To this end, the recent advance of machine learning (ML), owing to its ability to uncover hidden relationships from extensive data, offers new opportunities to accelerate the optimization of metal AM processes and improve our understanding of their processing-structure-property relationships. This thesis explores the application of ML techniques to improve the final quality of components fabricated through metal AM across various stages of the process. This thesis covers the following optimization aspects: (1) The individual influence of particle size distribution (PSD) on the powder flowability have been investigated. To reduce the time and effort required to characterize powder flowability, a reliable computer vision approach is established to evaluate powder flowability based on scanning electron microscope images. (2) ML technology is applied for parameter optimization by introducing a data-driven framework to establish process maps and employing optimization algorithms to suggest optimal processing parameters. (3) To facilitate efficient data analysis for synchrotron X-ray monitoring, various deep learning models are trained to identify and analyze the keyholes and generated pores from captured images. (4) To delve into the laser-metal interaction process, ML-predicted laser absorptance is integrated into a computational fluid dynamic model to accurately predict keyhole depth across various processing parameters. (5) A high-speed IR camera-based monitoring system is integrated into a customized LDED machine, and a comprehensive and reliable quality assessment metric is introduced for optimizing processing parameters for SS316 thin-wall samples. This thesis demonstrates that ML significantly improve the process efficiency and product quality in laser-based metal AM processes, offering a powerful tool for addressing many of the field’s most persistent challenges.
