A self-consistent hybrid global–2D model for SF 6 /Ar inductively coupled plasma etchers with neural-network surrogate acceleration
Abstract
We present a self-consistent hybrid global–2D model for reactor-scale simulation of SF 6 /Ar ICP etching. Unlike prior reactor-scale hybrids, which couple two spatial solvers, the framework couples a 0D global SF 6 /Ar chemistry solver to a 2D axisymmetric electromagnetic and species-transport solver through a shared masked-domain representation and a two-level Picard iteration. The 2D solver transports the nine neutrals with n e and T e , ion densities following from quasi-neutrality, while the full Lallement SF 6 /Ar chemistry and its coupled surface reactions are retained in the 0D model, at minutes of wall-clock time per operating point, far below a full multidimensional solve. The principal advance is that quantities ordinarily prescribed in reactor-scale 2D simulations, such as the power-coupling efficiency η, the electron-density profile n e (r,z), and the electron-temperature profile T e (r,z), are instead emergent outputs, so that the operating point determines the plasma state. Benchmarked against spatially resolved wafer-plane fluorine measurements, the model reproduces the absolute wafer-center fluorine density to within 6–20% on the calibration composition, over-predicts a blind composition by 1.5–2.1×, and under-predicts the measured center-to-edge [F] non-uniformity by roughly 15 percentage points. Both absolute-density residuals lie within the combined measurement and rate-coefficient uncertainty. A neural-network surrogate reproduces the wafer-relevant atomic-fluorine and SF 6 fields at sub-second inference, an 872× acceleration on a local workstation and 1750× on the NCSA Delta HPC system, each against the standalone chemistry–transport solve on that platform; a 21-channel extension emulating the reduced 2D state runs at 55–91×. An LXCat-based electron-kinetics analysis finds the dominant low-energy rates distribution-insensitive within roughly 20%, supporting the Maxwellian-averaged rates retained. Because the plasma state is computed rather than fitted, the converged model serves as a predictive instrument for testing physical hypotheses and operating scenarios. Together, the validated model and its surrogate form the predictive kernel of a reactor-scale SF 6 /Ar ICP digital twin for near-real-time recipe development.
