A Deep Learning Framework for Crater Identification and Classification Using Plume-Surface Interaction Experimental Data

Supersonic plume impinging on a granular surface during a Moon/Mars landing causes erosion, resulting in crater formation and ejecta flow. The multiphysics processes involved in the plume-surface interaction (PSI) under rarefied conditions make it challenging to predict and model the cratering and ejecta dynamics. The crater geometries formed during such interactions can provide insights into the underlying erosion mechanisms, aiding the development of predictive models and scaling laws for PSI in lunar/martian environments. A set of experiments called the Physics Focused Ground Test (PFGT) was conducted at NASA Marshall to investigate the PSI in rarefied conditions. This study develops an automated algorithm for extracting and classifying crater geometries from high-speed PFGT videos. The experiments span various mass flow rates, nozzle heights, ambient pressures, and regolith materials. We create a labeled dataset for training edge-based and segmentation-based deep neural networks (DNNs) to identify crater geometry. The segmentation-based DNN (which determines the crater region instead of just the edge) performs much better than the edge-based network. The extracted geometry from the trained network is then used as an input to a convolutional neural network to classify the crater shape. The trained classifier network can automatically categorize video frames into crater shapes: parabolic, annular, conical-parabolic, conical-annular, and conical-deep.