Assessment of Machine-Learning-Based Forecasting Models for Plume-Surface Interaction

Plume-surface interaction (PSI) is a complex multiphysics phenomenon due to the coupled interaction between high-speed jet flow, granular ejecta dynamics, and evolving surface erosion during lander descent. These interactions under rarefied conditions, relevant to Moon or Mars, can pose risks to landers and nearby equipment. Understanding these coupled phenomena is key to capturing the underlying PSI process. To better understand PSI-induced cratering, NASA conducted the Physics Focused Ground Test (PFGT) campaign, which captured high-speed videos of crater evolution across various mass flow rates, ambient/vacuum pressures, nozzle heights, and granular bed properties. These tests produced a range of crater shapes, such as annular, parabolic, and composite forms, exhibiting diverse temporal evolution behaviors.

This study employs neural-network-based time-series forecasting models, including Long Short-Term Memory (LSTM) networks and transformer architectures, to predict the temporal evolution of crater geometry, quantified by its volume, depth, and aspect ratio, across the test parameter space. The predictive accuracy and training data requirements of each model are assessed and compared. Where necessary, the PFGT dataset is supplemented with in-house experimental and simulation data to evaluate model performance under diverse operating conditions.