Development and Use of a Neural Network for Optimizing Output of Accelerator with Large Control-Parameter Space

Accelerator systems with a large control-parameter space can be difficult to optimize. We report ongoing work to develop and train a neural network to act as a high-fidelity surrogate model. This model is used to find optimal parameter settings for given performance metrics. The accelerator system used is the Neutralized Drift Compression Experiment-II (NDCX-II). It is a high intensity ion induction linac at Lawrence Berkeley National Laboratory used to accelerate, shape, and compress a bunch of He⁺ ions . It has a 1m long drift section between the final focusing solenoid and target that is filled with plasma to neutralize beam space-charge in the final stage of the pulse compression to enable higher intensity on-target. NDCX-II is capable of delivering 0.7J/cm² within a ~1mm diameter spot on-target by compressing and accelerating an initial ~600ns, 135KeV pulse to ~1ns, 1MeV on-target. Approximately 40 parameters are used to vary the system with detailed simulations used to guide machine tuning. A NN consisting of dense fully connected layers is trained using experimental and simulation data. Here we report on the fidelity of the NN and progress towards increasing delivered fluence along with further experimental applications.