Development of dynamic spawning and annihilation of nodes in a Graph Neural Network, evaluating the developed model against plasma processes that occur in Electric Propulsion thrusters for spacecraft

Electric Propulsion (EP) systems, used to manoeuvre satellites in space, have become commonplace in recent years, with many technologies creating plasma interactions with electromagnetic fields to expel mass at high velocities to propel the spacecraft. The increased specific impulse of EP systems over chemical propulsion systems can lead to enhanced overall net gains from a system perspective in given mission scenarios, making EP an attractive solution. A significant ongoing challenge when developing EP systems is the simulation, analysis, and qualification cycle that EP systems go through during their product development.

New EP systems can take years to develop due to the difficulty in determining and defining their performance and operation through simulation and physical testing. Simulation campaigns for some EP developments can take many months or even years to perform based on the complexity of the simulation model and the required accuracy of the simulated plasma process. Most models share a common feature in which they apply the Lorentz force equation and solve Maxwell’s equations to evolve a system in which plasma is present. Several methods exist for simulating different types of thrusters such as fully kinetic particle-in-cell (PIC) models, hybrid PIC models and fully fluid models. The fully kinetic model tends to have the best fidelity but requires significant computational power to be of use in 3D space. Fully fluid solvers are used for rapid and efficient plasma modelling over longer timescales but suffer from limitations such as an inability to model non-Maxwellian behaviour. Even then generalised models are not always able to simulate electric propulsion systems fully, due to some of the EP technological complexities, and so it is necessary for custom specific models to be created. Even if these are not important factors the computational resources for large scale models are not always available.

To solve many if not all these issues, a Graph Neural Network model dubbed PlasmaNet has been conceptually developed and through this studentship will be developed further. PlasmaNet takes inspiration from work carried out with Graph Nets for physics simulations by DeepMind. The model proposed will have three main components, the Encoder, the Processor and the Decoder. The encoder and decoder will be Multilayer Perceptron Neural Networks (MLPNN) and will compress the inputs into a latent space representation which will be passed into the processor and then decompress the low dimensional output of the processor back to high dimensional data respectively. At the current proposed state, the processor (graph net) will be performing node and edge predictions where these will be analogous to the change in a particle’s parameters and its relationship with other particles. The edge for each neuron will be drawn using the Debye length of the plasma. The graph will be (re)drawn with each new loop of the system.

However, one fundamental principle of plasma physics is not accounted for in this concept, species annihilation and creation. It is suggested that a traditional algorithm can be used for this after the predicted data is decoded but will likely be an inefficient methodology especially when many particles are simulated in turn not allowing for the full potential of using neural networks for this type of simulation. The aim of this study will be to investigate a way in which graph level predictions can be made where the goal is to design an algorithm to annihilate and spawn neurons dynamically.

This study will mainly focus machine learning and neural network where the plasma physics understanding required is minimal and can be learnt during the study. Simulation data from UCLA’s OSIRIS particle-in-cell (PIC) plasma simulation tool will be used to train the model. The successful completion of this project will see the underpinning principles application to a variety of fields aside from plasma physics.

The project would suit a graduate in Engineering, Physics or Applied Mathematics with interests in Computing. It is desirable for applicants to be familiar with Python, Fortran or C, machine learning (especially neural networks), Linux and high-performance computing.