Plasma simulations no longer wait out slow ion relaxation: similarity reduction plus a dual-branch model cuts convergence cycles by a median 27.7 times
Synopsis
Particle simulations of low-pressure capacitively coupled plasmas have long been limited by slow ion relaxation, often needing thousands of radio-frequency cycles to reach a periodic steady state. This work uses discharge similarity to reduce the three-dimensional control space of gas pressure, gap distance and driving frequency into a two-dimensional similarity space, then a dual-branch machine-learning model reconstructs the ion density profile and conditional ion velocity distribution as the simulation's initial condition. In a 1D3V argon discharge, this initialization achieves a median 27.7-fold and maximum 56.8-fold speedup over uniform initialization, and maps discharge-mode transitions and low-density jumps within the similarity space.
Interpretation
Kinetic simulations of low-pressure capacitively coupled plasmas no longer have to start from a uniform initial condition; the ion kinetic state can be predicted directly in similarity space and injected into the particle simulation as a near-steady-state initial condition. The conventional approach initializes with a uniform density profile and must evolve thousands of radio-frequency cycles before slow ion relaxation converges to a self-consistent periodic steady state. Initialized from this state, the cycles needed to reach the same convergence criterion give a median 27.7-fold and maximum 56.8-fold speedup over uniform initialization; the model saw only the base similarity plane in training, while the test set spans different scaling factors.
The method splits the three-dimensional control parameter space into a two-dimensional similarity space plus an independent scaling factor, so the model learns only the scale-invariant part. Different combinations of pressure, gap distance and driving frequency do not all represent independent discharge states; under synergistic scaling a weakly ionized Boltzmann-Poisson system can correspond to the same reduced state. In the 1D3V argon discharge, discharge voltage is fixed at 200 V and the secondary electron emission coefficient at 0.2, and the training set uniformly samples 81 independent operating conditions within the base similarity plane.
The dual-branch model handles the smooth macroscopic density profile and the local kinetic features in two-dimensional phase space separately. The steady-state ion distribution function is factorized into an analytically determined scaling factor, a base-scale density profile and a conditional ion velocity distribution function, with the latter two invariant along the scaling axis. The density branch uses principal component analysis plus a multilayer perceptron, the shape branch uses a convolutional network with deconvolution to reconstruct local probability distributions, and the two outputs are recombined with the analytic scaling factor into the full dimensioned ion velocity distribution function.
Perspective
The demonstrated scope is a 1D3V argon capacitively coupled discharge with discharge voltage fixed at 200 V and secondary electron emission coefficient 0.2, with training data generated by the in-house ASTRA electrostatic particle simulation code; cross-scale reliability was checked against independent simulations on parameter cross-sections at two corresponding scaling factors.
The similarity reduction relies on assumptions such as weak ionization and negligible neutral gas depletion, and scaling of the collision term requires collision energy to remain invariant under scaling; prediction error and speedup are statistical results over the test set and the parameter space.
