| dc.description.abstract | Retro-propulsive landings on extraterrestrial surfaces pose significant risks due to the interaction between the supersonic plume from rocket nozzles and the ground. The plume–surface interaction (PSI) releases dust and gravel, creating uncertainties in landing safety, spacecraft integrity, and environmental impact. To improve the understanding of PSI, the Physics Focused Ground Test (PFGT) campaign was conducted at NASA Marshall, generating high-quality PSI data on impingement pressures, plume flow structures, crater evolution, and ejecta dynamics under various nozzle heights, mass flow rates, ambient pressures, and regolith simulants. The crater geometries and their transitions reflect underlying erosion mechanisms such as viscous erosion, bearing capacity failure, and diffusion-driven shearing, which remain poorly understood under different flow and soil conditions.
This thesis aims to develop a predictive model for crater evolution by extracting and classifying crater geometries from PFGT high-speed experimental videos. Firstly, two different Convolutional Neural Networks (CNNs), edge-based and segmentation-based, were trained to automate crater edge extraction and region segmentation. The segmentation-based network outperformed the edge-based network, accurately identifying crater profiles across most experimental conditions. Secondly, another CNN, trained on manually labeled data, was employed to automate crater shape classification, which included parabolic, annular, conical-parabolic, conical-annular, and conical-deep categories, with high accuracy, thereby enabling the automated detection and analysis of crater shape transitions.
Finally, a data-driven parametric model is investigated using polynomial functions to model crater volume evolution under various operating conditions. Polynomial feature maps are employed to model crater volume evolution. Single-parameter coefficient models are used to model crater-volume evolution when only one operating parameter is varied. They provide a robust framework for modeling parameter-dependent temporal evolution; however, the limited experimental data restricted the predictive accuracy of the model, especially in transitional regimes. | en_US |
