Abstract Scope |
1. Introduction:
Producing defect-free, high-quality additive manufacturing (AM) parts in a time-efficient and cost-effective way is a major challenge to the materials community. Defects can be caused by complex mechanisms depending on a large range of process parameters and material properties. Currently, defects cannot be minimized by either calculations or high-cost experimental trials. There are no simple criteria to predict and reduce defects in additive manufacturing based on scientific principles. A recource is to synthesize machine learning, experimental testing, and mechanistic models to reveal the defect formation mechanisms and predict defect formation. In this research, we find an easy-to-use, verifiable balling susceptibility index using supervised machine learning algorithms to forecast and reduce the balling defect in laser assistant powder bed fusion (L-PBF).
2. Technical approach:
We investigate the influence of six important mechanistic variables on balling defect based on one hundred and sixty-six sets of independent experimental data from the literature on balling formation. The six mechanistic variables are volumetric energy density, surface tension force, Marangoni number, Richardson number, pool aspect ratio (pool length/pool depth), and solidification time of the pool, and they can capture the combined effects of process parameters and alloy properties. The data include six commonly used alloys, AlSi10Mg, Aluminum 357, stainless steel 316, Co-Cr, Ti6Al4V, and Inconel 718. A balling susceptibility index is derived from a genetic algorithm to predict balling defects based on these six mechanistic variables. In addition, the hierarchical importance of mechanistic variables on balling can be explained from the coefficients of the balling susceptibility index and three machine learning algorithms. The effect of process variables on balling defect formation is analyzed. The Marangoni number and solidification time are found to be the first and second most important variables for the formation of balling defects, respectively for the alloys.
3. Results and discussions:
We find an easy-to-use, verifiable balling susceptibility index that combines the effects of process parameters and alloy properties. The balling susceptibility index can predict balling defect with 90% accuracy using 166 cases. The six mechanistic variables reveal the mechanisms of balling defect based on scientific principles with machine learning. We provide six balling susceptibility maps for six commonly used alloys where the trends of the balling susceptibility index with the well-known L-PBF variables are consistent with the common industrial practice for both fusion welding and additive manufacturing.
4. Summary and conclusion:
In summary, experimental data sets, model results, and machine learning algorithms are combined to reveal the mechanisms of balling formation and predict appropriate conditions to prevent balling defects in additive manufacturing. The approach used here can help to reduce other common defects in AM such as cracking, porosity, lack of fusion, and surface roughness. Reduction of defects based on scientific principles and machine learning will improve part quality, reduce cost, and allow printing of new components. |