A dual machine learning framework for the solution of the inverse scattering problem in cylindrical plasmas

Abstract Machine learning algorithms are widely used to replace or accelerate high-fidelity codes and to provide physical insight in complex problems. In this work, a random forest is applied to replace a 3D full-wave simulation performed using the COMSOL Multiphysics software. In our simulation, a microwave beam is propagating through a cylindrically shaped plasma and is scattered by it. The training data of the random forest consist of the electron density profiles and the corresponding distributions of the microwave beam power after the interaction with the plasma. The random forest accurately reproduces the resulting scattered beam distribution for a given density profile. A synthetic dataset is then created with the forest, which in turn is used to train a neural network (NN). The NN is trained to solve the inverse problem, predicting a parametric description of the electron density profile for a given beam power profile. The NN is tested on new data generated with COMSOL and successfully predicts the electron plasma density used in the simulation.