Neural Network-Based Predictions of Crater Evolution During Plume Surface Interaction

Supersonic rocket exhaust plumes impinging on Lunar and Martian surfaces mobilize regolith and form craters whose evolution is strongly dependent on plume characteristics and ambient conditions. This study examines crater formation and temporal evolution in plume-surface interaction (PSI) using the data from Physics Focused Ground Test (PFGT) experiments conducted in NASA Marshall’s 15-ft vacuum chamber under various mass flow rate (\(\dot{m}\)), ambient pressure (\(p_{amb}\)), and nozzle height (\(h/d_e\)). Crater edges and their time evolution are extracted from high-speed videos using a U-Net–based convolutional neural network, enabling robust reconstruction of crater geometry. The resulting crater volume time evolution exhibits regime-dependent transitions, e.g., annular to parabolic, parabolic to conical–parabolic, etc., whose onset varies with operating conditions. Polynomial feature maps are employed to model crater volume evolution. Single-parameter coefficient-based models are first used to assess crater-volume scaling when only one operating parameter is varied, demonstrating reduced predictive accuracy in transitional regimes. A feedforward neural network with a polynomial feature map (PolyNet) is then used to predict crater-volume evolution on the full set of operating parameters. The model reproduces the crater volume evolution on the training operating conditions with an average relative \(L_2\) error of 0.02, but exhibits significantly larger errors on held-out testing conditions, with an average relative \(L_2\) error of 0.278. These results indicate that sparse sampling near regime transitions limits the predictive capability of PolyNet, highlighting the need for additional experimental data in transition regimes to robustly model and predict crater evolution.