Automatic detection of equiaxed dendrites using computer vision neural networks

Equaixed dendrites are frequently encountered in solidification. They typically form in large numbers, which makes their detection, localization, and tracking practically impossible for a human eye. In this paper, we show how recent progress in the field of machine learning can be leveraged to tackle this problem and we present computer vision neural network to automatically detect equiaxed dendrites. Our network is trained using phase-field simulation results, and proper data augmentation allows to perform the detection task in solidification conditions entirely different from those simulated for training. For example, here we show how they can successfully detect dendrites of various sizes in a microgravity solidification experiment. We discuss challenges in training such network along with our solutions for them, and compare the performance of neural network with traditional methods of shapes detection.