A joint research team from Shanghai Jiao Tong University employed synchrotron X-ray computed tomography (CT) combined with 3D reconstruction to classify gas-atomized powders into five categories—irregular, satellite, normal, hollow, and multi-defect—based on formation mechanisms. They extracted nine morphological parameters from 2,598 powder samples to train machine learning models. Through principal component and correlation analysis, four key distinguishing parameters were identified, achieving high classification accuracy across all models. An improved nozzle designed accordingly was evaluated using particle defect distribution (PDD), resulting in an approximately 300% increase in the mass fraction of high-sphericity normal powder. This approach establishes a closed-loop feedback path from defect morphology quantification to nozzle design, providing quantifiable data to support iterative design of gas-atomization powder production equipment.