Statistical and Physical Modeling of Inert Gas Atomization for Metal Powders Production

Gas atomization is a key process for producing metal powders with spherical morphology, suitable for Additive Manufacturing technologies. This study introduces a statistical approach for the development of a predictive model for estimating the median particle size (D₅₀) in close-coupled gas atomizers, comparing it with existing physical models. The statistical model is based on data collected from 67 inert gas atomizations performed in a VIGA system using argon as the atomizing gas. The effect of the process parameters was investigated in terms of gas temperature and pressure, melt temperature and flow rate. The Particle Size Distribution (PSD) was characterized both via laser diffraction analysis (Malvern Mastersizer) and image analysis (Camsizer). Preliminary data analysis identified the volume-based Gas-to-Metal Ratio (GMRv) as a stronger predictor of D₅₀ than mass-based Gas-to-Metal Ratio (GMR). A clear trend was observed between the increase in GMRv and the decrease in D₅₀ together with a consistent reduction in clustering related to the atomization gas temperature variation. The statistical model developed is a linear regression model incorporating metal mass flow rate and gas velocity as independent variables, achieving satisfactory fitting of the experimental dataset (R² = 0.85 Malvern, R² = 0.87 Camsizer). Preliminary analysis of the as-atomized PSD width (span) revealed a significant correlation with GMRv. However, applying the same set of independent variables used for D₅₀ did not yield a similarly satisfactory fit, highlighting the need for more advanced statistical methods. The narrow range of span values, in comparison to the ranges of the independent variables, may necessitate dataset normalization or the use of more robust regression models. The accuracy of existing semi-empirical models (Lubanska, Rao-Mehrotra, Dunkley) was evaluated considering the same dataset. Lubanska model showed limited applicability (R² = 0.66 Malvern, R² = 0.73 Camsizer), losing effectiveness when considering atomizing gas temperatures different from ambient values. Rao-Mehrotra’s modification of Lubanska model showed improved fitting of the experimental dataset (R² = 0.71 Malvern, R² = 0.81 Camsizer), together with the model proposed by Dunkley (R² = 0.72 Malvern, R² = 0.80 Camsizer), even when considering a wide range of gas temperatures. An optimized version of Dunkley’s model was proposed by converting the gas temperature fixed exponent of -0,5 into an additional adjustable parameter. Introducing a degree of freedom improved the fitting of the experimental data significantly, making the model more specific to the industrial atomizer considered in this study (R² = 0.83 Malvern, R² = 0.88 Camsizer). A comparative analysis based on 24 new atomization processes confirmed the highest predictive power for the statistical model (R² = 0.73 Malvern, R² = 0.64 Camsizer), although the optimized Dunkley’s model provides comparable results (R² = 0.70 Malvern, R² = 0.60 Camsizer). Further improvements to the model's predictive power will need to increase regression robustness through a rigorous repartition between training and testing data.