Abstract Scope |
Introduction:
Common defects in additive manufacturing affect the mechanical properties and serviceability of parts. For example, lack of fusion voids and solidification cracking are known to adversely affect tensile properties. The formation of balling defects affects the part quality and fatigue properties. The formation of these defects depends on multiple complex physical processes and there is no simple method to mitigate them based on scientific principles. Here, we identify conditions for minimizing the lack of fusion, cracking, and ball formation in additively manufactured metallic parts by using physics-informed machine learning and experimental data. In this approach, we identify several mechanistic variables that can capture the mechanisms of defect formation. In physics-informed machine learning, the physics of defect formation captured through the computed values of the mechanistic variables is augmented into machine learning.
Technical approach:
We collect independent experimental data on the three types of defect formation from the literature for stainless steel, nickel, titanium, and aluminum alloys. The dataset contains about a hundred data points for each defect which are used to train, validate, and test a neural network, a random forest, and a decision tree. First, the alloy properties and printing process parameters such as laser power, scanning speed, layer thickness, hatch spacing, preheat temperature, and substrate thickness are used to compute the mechanistic variables using a heat transfer and fluid flow model of the additive manufacturing process. The model solves the equations of conservations of mass, momentum, and energy in a 3-D discretized solution domain that includes deposit, substrate, and shielding gas. Thermophysical properties of stainless steel, nickel, titanium, and aluminum alloys required for the heat transfer and fluid flow model are computed by thermodynamic calculations using the commercial software JMatPro. The mechanistic variables include peak temperature, molten pool dimensions, cooling rate during solidification, solidification parameters, stresses originated due to solidification, Fourier number, and Marangoni number. Computed values of these mechanistic variables are used in neural networks to forecast the occurrence of defects. Then the hierarchical influences of these mechanistic variables on the defect formations are predicted using random forests and decision trees.
Results and discussions:
Temperature distribution, molten pool dimensions, cooling rate during solidification, solidification parameters, stresses originated due to solidification are important factors that control the lack of fusion voids, balling, and solidification cracking. The neural networks trained using the computed values of the mechanistic variables can predict all three types of defects with more than 95% accuracy. Stresses that are generated due to shrinkage during the solidification are found to be very important in controlling solidification cracking. In addition, both molten pool width and depth are found to significantly affect the lack of fusion void. Small pool exhibits improper fusional bonding with the neighboring tracks and results in a lack of fusion voids. Furthermore, the ratio of pool length to pool depth is highly responsible for ball formation. A very long pool may be unstable and may break into small pools that form balls on the part surface after solidification.
Summary and conclusions:
In summary, physics-informed machine learning which is a synergistic combination of heat transfer and fluid flow model of additive manufacturing and machine learning algorithms such as neural network, decision tree, and random forest along with experimental data are used to predict conditions to mitigate lack of fusion, cracking and the balling defects in additively manufactured parts. Predictions of defects using the mechanistic variables as inputs are more accurate than that using the raw additive manufacturing process parameters. Therefore, the computed mechanistic variables are superior to the raw printing process parameters in forecasting defect formation. The hierarchical influence of the variables on the defect formation can guide engineers to select the most important variables to tune in to reduce the common defects in additively manufactured parts. |