Speaker
Description
Atmospheric re-entry tools and their predictions are increasingly utilized in the satellite design process due to the demand for safe satellite disposal to minimize the rising debris population in orbit. Accurate predictions of re-entry, particularly in aerothermodynamics, are crucial to evaluate the probability of demise and casualty risk associated with re-entering vehicles. Destructive atmospheric re-entry is an inherently complex environment to model, with hypersonic re-entry eliciting shock waves, thermal and chemical non-equilibrium and high-temperature flows. Traditionally, aerothermodynamics are predicted using low- or high-fidelity methods. Low-fidelity methods, based on semi-empirical and analytical correlations are computationally efficient but disregard the complex physics of destructive re-entry. This leads to substantial uncertainties in the predicted fragmentation and trajectory of the vehicle. Conversely, high-fidelity methods such as Computational Fluid Dynamics (CFD), are increasingly capable of resolving the complex phenomena of re-entry, but at a significantly high computational cost. Achieving a balance between the accuracy and cost of re-entry predictions is crucial for the design for demise processes of satellites.
Emerging research explores a promising approach using Deep Learning techniques to create a synergy between low- and high-fidelity methods to establish the balance of accuracy and computational efficiency. Deep learning methods, such as the neural network architectures of feed-forward artificial neural networks (ANN) and convolutional neural networks (CNN) leverage a set of high-fidelity simulations to predict surface quantities distributions like heat-flux or pressure. These predictions incorporate the complex physics of numerical high-fidelity simulations to provide a superior accuracy when compared with low-fidelity or physically-blind methods. The computational cost of a given prediction by deep learning models is however computationally inexpensive, allowing for the probabilistic analysis of design for demise processes of satellites to be carried in a computationally efficient and accurate capacity.
The purpose of this study is to investigate the implementation and applicability of deep learning models for the prediction of aerothermodynamic loads during destructive hypersonic re-entry. Geometrically simple test cases, such as the ring and cube will be considered to analyse the training data demand of deep learning models, as well as their ability to resolve complex non-linearities of re-entry for rarefied and continuum flow. The Automated Transfer Vehicle (ATV) will also be used to investigate the deep learning models applicability to complex test cases in the continuum flow regime. The optimal neural network architecture and training procedure for these types of problems will be explored as part of this work. Python’s machine learning library TensorFlow will be used for the creation and deployment of the neural network models of this study. In addition, the multi-fidelity re-entry tool TITAN will be used for all time-propagate re-entry analysis and the high-fidelity tools SU2-NEMO and SPARTA are utilized for high-fidelity simulations in the continuum and rarefied flow regimes, respectively. This study will therefore demonstrate the potential feasibility, and subsequent impact during atmospheric re-entry simulations on satellite trajectory and demisability when deep learning neural network models are employed. The findings will be assessed in terms of the accuracy and computational efficiency delivered by these models.
