AI4Plasma I: Data-Efficient Surrogate Modeling of Capacitively Coupled Plasma Fields
Organizations: Theory Lab, Huawei Leibniz Research Center, Shatin, Hong Kong SAR, China. · Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Shatin, N.T. 999077, Hong Kong SAR, China · Department of Mathematics, Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong SAR, China · Beijing International Center for Mathematical Research, Peking University, Beijing 100871, China.
Abstract
Reference capacitively coupled plasma (CCP) simulations provide spatially resolved plasma fields but are costly to generate across operating conditions. We compare the data efficiency of five surrogate architectures--Coordinate MLP, PCA-Net, U-Net, Fourier neural operator (FNO), and Transolver--for predicting RF-cycle-averaged electron density, ion density, electric potential, and electron temperature in 1D planar and 2D axisymmetric GEC CCP benchmarks. We assess the number of reference simulations needed to achieve 1% and 2% mean relative L2 error for parameter-interior and parameter-boundary test cases, and examine architecture-specific mechanisms. FNO achieves the lowest or joint-lowest reference-case requirement in seven of eight comparisons. In 1D, FNO reaches 1% interior error with 50 training cases, compared with 100 for Transolver and 200 for U-Net. In 2D, FNO and U-Net each reach 1% interior error with 100 cases; FNO and Transolver each reach 2% parameter-boundary error with 200 cases, compared with 400 for U-Net. At the stricter 1% 2D parameter-boundary target, Transolver requires 400 cases versus 800 for FNO. Additional analyses investigate FNO spectral modes and padding, U-Net skip connections, Transolver adaptive slicing, and PCA-Net projection versus prediction error. The results support FNO as a data-efficient baseline for the tested CCP settings, while showing that the preferred architecture depends on the accuracy target and operating-parameter region.
Figures & tables
| Item | 1D planar CCP | 2D axisymmetric GEC CCP |
|---|---|---|
| Geometry | Parallel plates, | Axisymmetric GEC chamber |
| Operating conditions | – ; – ; – | Same parameter ranges |
| Spatial grid | 128 cells | ( ); 4624 active cells |
| Target fields | (RF-cycle averaged) | (RF-cycle averaged) |
| Valid reference cases | 3000 | 1000 |
| Train / interior / boundary |
| 1D | 2D | |||||||
| Interior | Boundary | Interior | Boundary | |||||
| Model | ||||||||
| Coordinate MLP | 200 | 800 | 400 | NR | 600 | NR | NR | NR |
| PCA-Net | 200 | 400 | 400 | 1600 | 400 | 600 | NR | NR |
| U-Net | 100 | 200 | 400 | 800 | 50 | 100 | 400 | 600 |
| FNO | 50 | 50 | 200 | 400 | 50 | 100 | 200 | 800 |
| Benchmark | Model | Interior (%) | Boundary (%) |
|---|---|---|---|
| 1D | Coordinate MLP | 0.718 [0.685,1.437] | 1.206 [1.149,2.339] |
| PCA-Net | 0.314 [0.274,0.334] | 0.706 [0.658,0.746] | |
| U-Net | 0.144 [0.140,0.146] | 0.501 [0.490,0.533] | |
| FNO | 0.183 [0.177,0.188] | 0.461 [0.453,0.470] | |
| Transolver | 0.180 [0.179,0.186] | 0.427 [0.394,0.457] | |
| 2D | Coordinate MLP | 1.990 [1.748,2.124] | 2.708 [2.620,2.847] |
| Application requirement | Preferred model | Supporting evidence |
|---|---|---|
| Limited reference data | FNO | Lowest or joint-lowest reference-data requirement in seven of eight 1%/2% target comparisons. |
| High-accuracy interior-field prediction | U-Net | Lowest interior errors at the largest tested training-set sizes: 0.144% (1D) and 0.126% (2D); reaches the 2D interior 0.5% target with 200 cases. |
| Stringent 2D parameter-boundary prediction | Transolver | Reaches the 1% target with 400 rather than 800 cases for FNO, saving approximately 1521 CPU core-hours of reference generation. |
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Item | 1D planar CCP | 2D axisymmetric GEC CCP |
|---|---|---|
| Geometry and grid | gap; 128 cells | ; powered radius ; ( ) grid; 4624 active cells |
| Gas and species | Ar; ; electrons and positive ions | Ar; ; electrons and positive ions; no metastable branch |
| Reactions and energy | Tabulated direct ionization and excitation | Tabulated direct ionization and excitation; stepwise, superelastic, quenching and pooling channels disabled; ionization–excitation–elastic energy-loss model |
| Transport | Pressure-scaled constant mobility/diffusion | Pressure-scaled constant mobility/diffusion; electron-moment transport table |
| Wall treatment | Constant electron wall flux; | Sakiyama-type electron/ion particle fluxes [ 70 ] ; ; electron-energy wall flux uses adjacent-cell ; fixed vessel-wall potential ( ) |
| Discretization and Poisson | Finite volume; upwind flux; semi-implicit update | Finite volume; upwind flux; semi-implicit Poisson update with electron-drift coupling; SparseLU, tolerance |
| System | Case | Comparison | ||||
|---|---|---|---|---|---|---|
| 1D | low | 200–400 steps/RF | 0.938 | 0.780 | 0.392 | 3.572 |
| 1D | center | 200–400 steps/RF | 0.237 | 0.230 | 1.069 | 5.721 |
| 1D | high | 200–400 steps/RF | 0.140 | 0.144 | 1.663 | 0.816 |
| 1D | center | 128–256 cells | 4.254 | 4.206 | 4.587 | 13.815 |
| 1D | center | 256–512 cells | 1.996 | 1.972 | 1.874 | 5.871 |
| 2D | low | 150–300 steps/RF | 0.348 | 0.391 | 1.193 | 0.919 |
| Architecture | Retained 1D candidates | Retained 2D candidates |
|---|---|---|
| Coordinate MLP | width ; depth ; epochs ; LR | width ; depth 4; epochs 240; LR ; constant scheduler |
| PCA-Net | rank ; coefficient width ; epochs ; LR | rank ; coefficient width 128; epochs 240; LR ; constant scheduler |
| U-Net | base width ; depth ; cap ; epochs ; LR ; constant/OneCycle pilots | base width ; depth at width 64; cap 256; epochs 1000; LR ; OneCycleLR |
| FNO | corrected branch: width ; 3 layers; modes ; padding ; constant-900/cosine-tail-200 | width 64; 4 layers; modes ; padding variants F0–F3; epochs 1000; LR ; constant/OneCycleLR |
| Transolver | hidden width ; layers ; heads ; slices ; epochs ; LR ; constant/OneCycle pilots | hidden width 128; 4 layers; 4 heads; slices ; epochs 1000; LR ; OneCycleLR |
| System | Model | Architecture | Epochs | LR | Batch | Scheduler |
|---|---|---|---|---|---|---|
| 1D | Coordinate MLP | Width 128; 3 hidden layers | 240 | 0.001 | 32 | constant |
| 1D | PCA-Net | Rank 32; coefficient width 192 | 240 | 0.0005 | 32 | constant |
| 1D | U-Net | Base 32; depth 3; cap 256 | 600 | 0.001 | 32 | OneCycleLR |
| 1D | FNO | Width 128; 3 layers; 64 modes; padding 0.10 | 200 | 0.001 | 32 | cosine-tail |
| 1D | Transolver | Width 192; 4 layers; 8 heads; 16 slices | 600 | 0.0005 | 16 | OneCycleLR |
| 2D | Coordinate MLP | Width 256; 4 hidden layers | 240 | 0.0005 | 16 | constant |
| System | Model | Runs | Median h/run | Summed GPU-h |
|---|---|---|---|---|
| 1D | Coordinate MLP | 35 | 0.0024 | 0.1502 |
| 1D | FNO | 35 | 0.0083 | 0.5564 |
| 1D | PCA-Net | 35 | 0.0019 | 0.1236 |
| 1D | Transolver | 35 | 0.1157 | 7.3290 |
| 1D | U-Net | 35 | 0.0335 | 2.1635 |
| 2D | Coordinate MLP | 35 | 0.1606 | 7.9337 |
| 1D | 2D | |||
|---|---|---|---|---|
| Model | Interior | Boundary | Interior | Boundary |
| Coordinate MLP | NR | NR | NR | NR |
| PCA-Net | 800 | NR | NR | NR |
| U-Net | 400 | NR | 200 | NR |
| FNO | 200 | 2400 | NR | NR |
| Transolver | 400 | 2400 | 400 | NR |
| Field | Bulk | Near-axis | Transition | Outer radial |
|---|---|---|---|---|
| 0.393 | 0.391 | 0.478 | 0.358 | |
| 0.396 | 0.477 | 0.503 | 0.313 | |
| 0.293 | 0.353 | 0.332 | 0.187 | |
| 0.389 | 0.498 | 0.284 | 0.193 |
| Quantity | Learned slicing | Fixed radial slicing |
|---|---|---|
| Radial support RMS (cm) | 2.591 | 0.292 |
| Centroid pairwise distance (cm) | 1.113 | 3.445 |
| Effective occupied slices | 3.747 | 7.913 |
| Field | Test set | Projection (%) | Prediction (%) | Ratio |
|---|---|---|---|---|
| interior | 0.09143 | 1.1586 | 13.34 | |
| boundary | 0.23239 | 2.2033 | 10.11 | |
| interior | 0.08196 | 1.0014 | 12.77 | |
| boundary | 0.22337 | 2.0164 | 9.78 | |
| interior | 0.01813 | 0.2500 | 13.53 | |
| boundary | 0.04560 | 0.4798 | 11.12 |
Neural Surrogate Modeling
Neural surrogate modeling uses learned models to approximate costly simulations, enabling faster prediction, design, control, and inference.
See all 469 papers