Data-Driven Based Prediction and Optimization of Balling Levels in Laser Powder Bed Fusion Additive Manufacturing.

Laser powder bed fusion has been demonstrated as a promising additive manufacturing technology due to its unique advantages, such as weight reduction, the ability to produce arbitrarily complex geometries and single-step manufacturing. However, the production quality may deteriorate due to the poor surface quality of deposited layers caused by the occurrence of the balling phenomenon, which hampers its widespread application. In this work, a data-driven framework is proposed to optimize the process parameters of laser powder bed fusion to achieve satisfactory balling levels. The effects of key process parameters on balling levels are also investigated. Specifically, an image segmentation-based method is introduced to quantitatively evaluate the balling levels on the interlayer surfaces of as-built specimens under various process parameter combinations. Considering the limited amount of experimental data, different machine learning models, including polynomial regression, support vector regression, and backpropagation neural networks, are developed to predict the balling levels within a predefined process parameter space. The predicted values from the best-performing model are then used as fitness values of individuals in an improved genetic algorithm to search for globally optimal process parameters. The final validation experiments confirm that the as-built parts fabricated using the optimized process parameters exhibit minimal balling levels, demonstrating the effectiveness and feasibility of the proposed framework for balling level prediction and optimization. This study provides valuable insights and practical guidance for enhancing the quality of specimens produced in the laser powder bed fusion process.