Plasma Operational Simulation (POPSIM): A Control-Oriented Simulation Toolbox for Parallel Simulation, System Identification, and Optimization

POPSIM is a control-oriented simulation toolbox developed using the machine learning framework Jax [1] to meet the simulation requirements for SPARC scenarios and control: rapid simulations performed between pulses in operations, automated qualification of off normal event detection, rapid adaptation of models to incoming data during commissioning, and the optimization of trajectories and controllers. New machine learning frameworks such as Jax now enable the development of models that span the full spectrum from highly structured (e.g. physics models with free parameters) to highly unstructured (e.g. neural networks). Such models can now enjoy the same properties traditionally reserved for neural networks including compilation, parallelism on GPU, and automatic differentiation. To leverage these new capabilities, we developed POPSIM for: 1) fast interactive simulation, 2) parallel simulation of different physics scenarios and off normal events, 3) identification of nonlinear models from data, and 4) optimization of scenarios, trajectories, and controllers. We report on progress integrating POPSIM and the SPARC plasma control system within the MOSAIC framework [2], and results calibrating POPSIM models to data from existing machines.

[1] Bradbury, James, et al. "JAX: composable transformations of Python+ NumPy programs." (2018).

[2] Teplukhina, Anna, et al. "Full-pulse simulations for SPARC using MOSAIC." APS Division of Plasma Physics Meeting Abstracts. Vol. 2023. 2023.