Towards Fast, Accurate Predictions of RF Power Deposition/Current Profile via Data-driven Modeling

Modeling of radio frequency (RF) actuators requires minutes of computation time to simulate a single time slice, however results on the ms timescale are desired for real-time experimental control and integrated modeling applications.  This work explores the use of new methods based on machine learning (ML) for the purpose of accelerating these computations through fast surrogate models.  Latin hypercube sampling methods ensure that the database of 16,000+ GENRAY/CQL3D simulations covers the range of 9 input parameters (n_e0, T_e0, I_p, B_φ, R₀, n_||, Z_eff, V_loop, P_LHCD) with sufficient density in all regions of parameter space.  A comparison of speed and accuracy for several ML regression methods (random forest, multi-layer perceptron neural network, Gaussian process) highlights strengths for each approach in creating a fast surrogate model.  Other RF machine learning topics include full-wave preconditioners, the inverse problem for RF current drive (i.e. mapping from a desired current profile to a set of plasma/RF parameters that will produce the desired current profile), and a Hamiltonian physics informed neural network to accelerate ray tracing calculations.