Learned predictions of cross-field transport from machine state in LAPD mirror configurations

Mirror machines are a simple and flexible fusion reactor candidate. This simplicity and flexibility provides opportunities for collecting diverse datasets for machine learning-guided exploration. As a proof-of-principle, we seek to learn models of transport mirror configurations to guide future experiments in the Large Plasma Device (LAPD). Measurements of density, temperature and cross-field particle transport in a variety of single-well magnetic mirror configurations were performed in the recently-upgraded LAPD. The new lanthanum hexaboride (LaB6) source allows access to higher densities and temperatures in the LAPD compared to the former barium oxide (BaO). The flexible magnetic geometry of the LAPD was utilized to cover a range of mirror ratios from 1 (flat) to 6. Discharge parameters such as the discharge voltage and gas puffing flow rates were also varied. Qualitatively, mirror configurations in LaB6 LAPD plasmas showed very sharp axial density gradients unseen in BaO plasmas with similar field profiles. Some of these configurations were randomly generated using Latin hypercube sampling to efficiently cover parameter space and enhance dataset diversity. A neural network-based machine learning model was then trained on data from these generated configurations to predict profiles and particle flux from machine state information and auxiliary diagnostics. A generative model was also trained on these data, and model performance will be benchmarked on previous, less diverse data. An overview of the LAPD mirror experiments and reconstruction and prediction accuracies will be presented