Two Dimensional TM010 Axially Symmetric RF Cavity Solver Based on Convolutional Neural Network

A convolutional neural network model to solve some quantities of two dimensional axially symmetric eigenmode simulations used for particle accelerators is proposed. The model is trained using simulated data using SUPERFISH eigenmode solver. The input is a randomly varied pillbox cavity image that is down-converted using first order interpolation method to reduce input size, while the output are the TM
010
frequency, axial electric field at $r=0$ for several axial points, and the scale factor used to normalize the electromagnetic energy. The model exhibits a good accuracy during training with less than 1% error, and could predict the relation between cavity radius and RF frequency for TM
010
mode. For a cavity geometry not similar to training samples, such as elliptical cavities, the frequency error increased to almost 3% while the total error is about 8%. However, the simulation duration is significantly reduced compared to simulation using SUPERFISH, even when using suboptimal hardware. Thus the model might be useful for cavity geometry optimization using parameter sweep.