A hybrid experimental-computational approach for predicting Ti-6Al-4V powder degree of spheroidization using artificial neural networks

A hybrid experimental-computational method was designed to predict the spheroidicity of Ti-6Al-4V powder processed through the radio frequency plasma spheroidization process. Twenty-three experimental runs were conducted to measure particle spheroidicity using optical microscopy. A validated computational fluid dynamics model developed in Ansys Fluent was then used to expand the dataset to 67 samples by simulating additional parameter combinations and varying particle size, plasma power, gas flow rate, and powder feed rate. The combined dataset was used to train a feedforward neural network in PyTorch, which showed improved performance with larger training sets. Sensitivity analysis and three-dimensional response surfaces revealed optimal process conditions (12-15 kW power for 60-100 µm powders; 0.6-1.0 kg/h at 40-60 slpm gas flow) to maximize spheroidicity. The hybrid approach proved reliable for predicting spheroidicity and offers actionable guidance for process optimization.