Feed-forward neural network surrogate model for rapid simulations of a high-temperature superconducting bulk undulator

This work presents the use of a feed-forward neural network (FFNN) surrogate model to significantly increase the speed of simulation time for a meter-long high-temperature superconducting bulk undulator. We generate a dataset of the undulator with bulks of different critical current densities using the finite element method (FEM), which is then used to train an FFNN. We show that the FFNN can output the results of interest of the full finite element model to an accuracy of 0.28%, while requiring a computation time of 200 ms instead of 5.7 hours with FEM. Finally, we use the FFNN to develop an inverse analysis to estimate the critical current density of each bulk from a given undulator field and show that this procedure can reproduce the critical current density to within 0.47% deviation in 12 iterations and 2.5 seconds. If the full FEM simulations were used, the computation times would be ~68 hours, demonstrating a speed increase of nearly 100 000.