Machine-learned trends in mirror configurations in the Large Plasma Device

Abstract

This study demonstrates the efficacy of ML-based trend inference using data from the Large Plasma Device (LAPD). The LAPD is a flexible basic plasma science device with a high discharge repetition rate (0.25-1 Hz) and reproducible plasmas capable of collecting high-spatial-resolution probe measurements. A diverse dataset is collected through random sampling of LAPD operational parameters, including the magnetic field strength and profile, fueling settings, and the discharge voltage. NN ensembles with uncertainty quantification are trained to predict time-averaged ion saturation current ($I_{\text{sat}}$ -- proportional to density and the square root of electron temperature) at any position within the dataset domain. Model-inferred trends, such as the effects of introducing mirrors or changing the discharge voltage, are consistent with current understanding. In addition, axial variation is optimized via comprehensive search over $I_{\text{sat}}$ predictions. Experimental validation of these optimized machine parameters demonstrate qualitative agreement, with quantitative differences attributable to Langmuir probe variation and cathode conditions. This investigation demonstrates, using ML techniques, a new way of extracting insight from experiments and novel optimization of plasmas. The code and data used in this study are made freely available.

arXiv comment16 pages, 11 figures

References in corpus (6)

• On Calibration of Modern Neural Networks
• Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles
• 2022 Review of Data-Driven Plasma Science
• On the Pitfalls of Heteroscedastic Uncertainty Estimation with Probabilistic Neural Networks
• EAST discharge prediction without integrating simulation results
• Full Shot Predictions for the DIII-D Tokamak via Deep Recurrent Networks