Unmet Need
Research and applications in science and engineering depend on a wide range of data. Recent advances in machine learning are particularly promising to provide feedback, optimization and instrument control in real-time during complex experimental processes. Instance segmentation, the process of correctly detecting and accurately identifying the relevant objects within an image, is a critical step in the advanced analysis of image data but is often complicated by the lack of appropriate quality and quantity of application-specific training data. Therefore, there is a strong need for the development of machine learning (ML) system for the optimization of floating zone furnace melt geometry to improve the accuracy of instance segmentation.
Technology Overview
Johns Hopkins researchers have demonstrated that machine learning methods, specifically deep neural networks, can be used to efficiently, and in real time, segment video frames during growth, identify the size and shape of the molten zone, and classify its stability. Transfer learning enables rapid, effective model training to discriminate between three common zone stability modes with under 1000 manually labeled training artifacts. Sustained drops in model confidence effectively indicate transitions between modes and enable operator intervention to restabilize the growth. In addition, Hopkins scientists have also found that this model, trained on labeled images of a single material family, effectively identifies zone transitions during growth of a second, distinct class of materials, i.e., ML/ Artificial Intelligence (AI) methods, avoids the need for per-material training. This automated and robust instance segmentation paves the way for the efficient acceleration and automation of the synthesis of new, functional materials by the widely used floating-zone method. It also provides the basis for techniques to extract chemically meaningful information from optical imaging data.
Stage of Development
Working prototype has been developed.
Publications
N/A
