The possibility to predict the 3D shape of air bubbles in air-water bubbly flows from 2D imagery contributes to accurate information on mass transfer properties. Achieving this goal is reached using Convolutional Neural Networks (CNNs). Trained networks accurately convert 2D images of single gas bubbles into the 3D shapes. The 3D shape output is represented as a weighted sum of spherical harmonic functions. Spherical harmonics are used for grid independence and dimensionality reduction.

The CNN input was generated by recreating a 20x20x100 cm gas-liquid bubble column in 3D software Blender. With Blender, images could be rendered for each of the 10.000 3D bubbles acquired from a front-tracking dataset. When comparing the renders with reference images an invisible edge around the bubble was discovered in both sets. The edge became visible when bubbles overlapped, or the lighting was dimmed. The cause is light reflection and refraction and was validated experimentally. This proof increases the validity of the input images used for training the CNNs. The results of the CNNs are accurate but not suitable for real-life applications yet. Before application, the training database should be more varied and noise handling must be improved.