AI4Plasma II: Geometry-Aware Surrogate Modeling for Variable Electrode Gap GEC Plasmas
Authors: Hao Yang, Zhenyu Zou, Wenshuai Jie, Luowei Yin, Yufeng Wei, Daniel Zhengyu Huang, Chun Chen, Kangxin Liu, Rubing Han, Haoran Guan, Yipei Chen, Peiting You, Yuwei Fan
Organizations: Theory Lab, Huawei Leibniz Research Center, Shatin, Hong Kong SAR, People’s Republic of China · School of Aeronautic Science and Engineering, Beihang University, Beijing 100191, China · Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Shatin, N.T. 999077, Hong Kong SAR, China · School of Mathematical Sciences, Peking University, Beijing 100871, 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
Industrial plasma process development requires broad exploration of operating and geometric parameters. Geometric parameters, particularly electrode spacing, affect discharge structure and processing uniformity. Exploring these parameter spaces through repeated high-fidelity simulations can be computationally expensive. Surrogate models can offer rapid predictions, but geometric variation complicates their application because the computational domain and discretization may change across samples. In this work, six geometry-aware surrogate models based on prescribed and learned geometry mappings, extended-domain representation, continuous coordinates, and geometry-native discretizations are compared in terms of their data requirements for accurate predictions of full plasma fields and ion density profiles near the wafer. The simulation dataset spans pressure, voltage, frequency, and electrode gap, with each reference case requiring an average simulation cost of 5.32 CPU h. To achieve a 2% mean error across the four plasma fields, prescribed-mapping Geo-FNO and Transolver require 100 training cases, compared with 250-1000 for the other models. Predicting ion density profiles near the wafer is more demanding and reveals larger differences among models, with prescribed-mapping Geo-FNO requiring the fewest training cases for the profile accuracy targets considered. To further reduce the data required for accurate predictions near the wafer, we introduce charge-separation weighting into the training loss, which reduces the prescribed-mapping Geo-FNO's minimum tested training requirement for a 5% profile error from 1000 to 100 cases. These results support prescribed-mapping Geo-FNO as a practical starting point for the variable-gap problem studied here when a suitable geometry mapping is available, while Transolver achieves the highest prediction accuracy with sufficient training data.
Figures & tables
Figure 1: Two-dimensional axisymmetric computational domain and mesh of the variable-gap GEC configuration. The electrodes are separated from the surrounding grounded wall by thin annular insulators.
Figure 2: Ion-density distributions at electrode gaps of (a) 2, (b) 4, and (c) 6 cm, with p=0.316Torr , VRF=200V , and fRF=40.68MHz . All panels share the same color scale.
Figure 3: (a)–(c) Axial plasma profiles at p=0.316Torr , VRF=200V , and fRF=40.68MHz . (d)–(f) Discharge averages versus electrode gap at different operating conditions. Changes in the discharge-averaged ion density and potential are more pronounced at small gaps. Increasing the gap lowers the mean electron temperature, particularly at low pressure.
Figure 4: (a) Radial profile locations. (b), (c) Midplane and wafer-adjacent ion-density profiles at the same operating parameters as in Fig. 3 . (d), (e) Radial nonuniformity versus electrode gap at different operating conditions. Increasing the gap generally enhances midplane nonuniformity, while its effect near the wafer is weaker and more dependent on the operating conditions.
Parameter
Symbol
Interval
Unit
Gas pressure
p
[0.10, 1.00]
Torr
RF voltage amplitude
VRF
[100, 300]
V
Driving frequency
fRF
[13.56, 67.8]
MHz
Electrode separation
g
[2.0, 6.0]
cm
Table 1: Operating-parameter ranges used for dataset generation.
Figure 5: Overview of the geometry representations and corresponding surrogate architectures considered in this study.
Model
ϵEe (%)
ϵni (%)
ϵne (%)
ϵϕ (%)
Mean (%)
Geometry mapping
Geo-FNO-A
0.848±0.028
0.864±0.027
0.859±0.015
1.046±0.029
0.904±0.014
Geo-FNO-L
0.961±0.024
1.014±0.015
0.981±0.020
1.082±0.051
1.010±0.022
Extended-domain representation
eDAFNO
1.096±0.118
1.166±0.101
1.080±0.089
1.940±0.069
1.320±0.092
Continuous-coordinate representation
Table 2: Four-field relative L2 errors on the original solver grids at N=1000 .
Figure 6: Data-efficiency comparison of the six surrogate models. Curves and shaded bands denote the mean and one sample standard deviation across five seeds. The horizontal dashed lines mark four-field mean-error levels of 1% , 2% , and 3% .
Model
N3%
N2%
N1%
Geo-FNO-A
50
100
1000
Geo-FNO-L
250
250
NR
eDAFNO
100
250
NR
C-MLP
500
1000
NR
PCNO
100
250
1000
Transolver
50
100
1000
Table 3: Minimum tested training-set sizes required to achieve 3% , 2% , and 1% mean four-field relative L2 error. Thresholds are determined from the mean errors across five random seeds. NR indicates that the target is not reached for N≤1000 . The smallest value in each column is highlighted in bold.
Figure 7: Physical regions used for error evaluation: (a) schematic partition of the GEC chamber and (b) region definitions.
Figure 8: Regional four-field relative L2 errors versus training-set size. Curves and shaded bands denote the mean and one sample standard deviation across five seeds, respectively, using the same 100 test cases at all training-set sizes. The sheath error remains higher than the bulk error as the training set grows.
Figure 9: Ion-density prediction errors at N=1000 versus the distance d=z−zb(g) from the lower electrode: (a) radial-profile relative L2 error, (b) radial-mean relative error, and (c) absolute nonuniformity error. Curves and shaded bands denote the mean and one sample standard deviation across five seeds, respectively. The gray dashed line in panel (a) shows the radial-mean charge-separation indicator ⟨χ⟩r of the reference solutions.
Axial midplane
Wafer-adjacent
Model
ϵprof (%)
ϵmean (%)
ϵC (pp)
ϵprof (%)
ϵmean (%)
ϵC (pp)
Geo-FNO-A
0.616
0.518
0.141
3.939
3.065
0.971
Geo-FNO-L
0.766
0.647
0.140
7.148
5.668
2.154
eDAFNO
0.672
0.531
0.148
13.804
12.511
2.924
C-MLP
0.729
0.528
0.189
15.288
13.869
3.417
PCNO
0.597
0.465
0.119
8.003
5.739
3.019
Table 4: Mean ion-density prediction errors at the axial midplane and wafer-adjacent layer for N=1000 . Values are averaged over 100 test cases and five seeds.
Figure 10: Wafer-adjacent ion-density radial-profile error versus training-set size. Curves and bands denote the mean and one sample standard deviation across five seeds. The horizontal dashed lines mark errors of 5% and 7.5% .
Model
N7.5%
N5%
Geo-FNO-A
100
1000
Geo-FNO-L
1000
NR
eDAFNO
NR
NR
C-MLP
NR
NR
PCNO
NR
NR
Transolver
500
1000
Table 5: Minimum tested training-set sizes required to achieve 7.5% and 5% wafer-adjacent ion-density radial-profile relative L2 error. Thresholds are determined from the mean errors across five random seeds. NR indicates that the target is not reached for N≤1000 . The smallest value in each column is highlighted in bold.
N=100
N=1000
Model
Four-field error (%)
Wafer-adjacent ni profile error (%)
Four-field error (%)
Wafer-adjacent ni profile error (%)
Geo-FNO-A
1.601→1.575
6.094→4.152
0.904→0.882
3.939→2.833
C-MLP
5.138→5.353
43.469→38.920
1.644→1.737
15.288→8.598
PCNO
2.185→2.196
18.551→12.172
0.899→0.867
8.003→4.394
Transolver
1.849→1.785
9.335→5.901
0.850→0.858
2.760→2.094
Table 6: Effect of the charge-separation-weighted training loss on prediction accuracy at N=100 and N=1000 . Each entry reports the error for λ=0→λ=4 , averaged over 100 test cases and five paired random seeds.
Accuracy target
Recommended model
Supporting evidence
Four-field mean error ≤2%
Geo-FNO-A
Reaches the target with N=100 , as does Transolver, but yields a lower mean error ( 1.601% versus 1.849% ).
Four-field mean error ≤1%
Transolver
Reaches the target with N=1000 , as do Geo-FNO-A and PCNO, and yields the lowest mean error ( 0.850% ).
Wafer-adjacent profile error ≤7.5%
Geo-FNO-A
Reaches the target with N=100 , compared with N=500 for Transolver.
Wafer-adjacent profile error ≤5%
Transolver
Reaches the target with N=1000 , as does Geo-FNO-A, but yields a lower profile error ( 2.760% versus 3.939% ).
Table 7: Recommended surrogate models for different accuracy targets in the variable-gap GEC benchmark. Recommendations use the original training loss.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Architectural configuration
Geometry mapping
Geo-FNO-A
Residual FNO; 4 Fourier blocks; feature width 36; 16 axial and 8 radial Fourier modes; padding (8,4) ; group normalization; GELU activation; trainable parameters: 1.334 M.
Geo-FNO-L
Geo-FNO with learned IPHI mapping; latent grid 130×40 ; 2 geometric and 3 standard Fourier layers; feature width 32; 16 axial and 8 radial Fourier modes; GELU activation; trainable parameters: 1.349 M.
Extended-domain representation
eDAFNO
Uniform grid: 250×50 ; lifting input [μ,r,z] ; mβ incorporated into every eDAFNO block; β=100 ; 4 explicit eDAFNO blocks; latent width 32; 16 axial and 8 radial Fourier modes; padding (48,4) ; GELU activation; trainable parameters: 1.054 M.
Coordinate-based continuous representation
Appendix
Table A1: Architectural configurations of the six surrogate models.
Model
Four-field error (%)
Wafer-adjacent ni error (%)
Original → Prescribed mapping
Change
Original → Prescribed mapping
Change
eDAFNO
1.32→0.97
−26.7%
13.80→9.23
−33.1%
C-MLP
1.64→1.88
+14.1%
15.29→16.45
+7.6%
PCNO
0.90→0.92
+1.8%
8.00→8.82
+10.3%
Transolver
0.85→0.94
+10.1%
2.76→3.15
+14.1%
Appendix
Table B1: Effect of applying the prescribed geometry mapping.
Model
Four-field error (%)
Wafer-adjacent ni error (%)
Original → Without gap input
Change
Original → Without gap input
Change
Geo-FNO-A
0.90→7.31
+708.2%
3.94→8.92
+126.5%
Geo-FNO-L
1.01→1.01
−0.1%
7.15→6.16
−13.9%
eDAFNO
1.32→1.31
−1.0%
13.80→13.07
−5.3%
C-MLP
1.64→23.35
+1320.3%
15.29→431.66
+2723.6%
PCNO
0.90→0.92
+2.6%
8.00→7.69
−3.9%
Appendix
Table B2: Effect of removing the explicit electrode-gap input.
Model
Four-field error (%)
Wafer-adjacent ni error (%)
Original → Reduced resolution
Change
Original → Reduced resolution
Change
Geo-FNO-A
0.90→1.83
+102.0%
3.94→5.49
+39.3%
Geo-FNO-L
1.01→1.08
+7.4%
7.15→6.45
−9.7%
eDAFNO
1.32→2.12
+60.9%
13.80→21.87
+58.4%
C-MLP
1.64→1.64
0.0%
15.29→12.39
−19.0%
PCNO
0.90→10.05
+1018.2%
8.00→47.91
+498.7%
Appendix
Table B3: Effect of reducing the spatial resolution.
Figure C1: Effects of learned slicing and token attention in Transolver at N=100 and N=1000 . The complete model is compared with two variants: one replaces learned slicing with a fixed 4×4 partition, and the other removes token attention. (a) Full-domain four-field error and (b) wafer-adjacent ion-density error. Percentages denote the relative error changes from the complete model at each training-set size.
Figure C2: Effects of retained Fourier modes and domain padding in Geo-FNO-A at N=1000 . The complete configuration uses 16×8 modes and 8×4 axial and radial padding. Panels (a,b) compare 8×4 , 16×8 , and 24×12 modes with approximately matched parameter counts. Panels (c,d) compare 8×4 , 8×0 , and 0×0 padding with fixed retained modes. The left and right columns report the full-domain four-field and wafer-adjacent ion-density errors, respectively. Percentages indicate changes relative to the complete model.
Figure C3: Effects of operator branches and Fourier quadrature in PCNO at N=1000 . The complete model is compared with variants trained without the local gradient branch, without the Fourier integral branch, or with uniform quadrature weights in the Fourier integral operator. Panels (a) and (b) report the full-domain four-field and wafer-adjacent ion-density profile errors, respectively. Percentages indicate relative changes from the complete model’s mean error.
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.
Zhenyu Zou, Junjue Liao, Hao Yang +11
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 +1
Fast and reliable plasma equilibrium prediction is essential for real-time tokamak operation and control, but conventional Grad-Shafranov (GS) solvers are often too costly for real-time deployment. We develop an AI surrogate framework and benchmark five architectures (MLP, CNN, FNO, Transformer, and KAN) on a numerical GS database with 100,000 IID and 10,000 OOD samples. Under a unified protocol, we evaluate accuracy, inference efficiency, model scaling, and robustness. We also establish device-level validation on the EXL-50U tokamak by linking numerical GS solutions, surrogate predictions, and the standard Shape Editor reference to assess simulation-to-device consistency. The surrogates achieve errors of 10−3-10−2 relative to GS solutions, while the GS-to-device discrepancy remains at 10−3. Transformer gives the best IID accuracy, whereas CNN offers the best balance of accuracy, robustness, and speed, reaching 0.7 ms TensorRT latency. On unseen plasma geometries and parameter regimes, CNN and FNO show the strongest extrapolation stability, with 4%-5% relative L2 error, while models with weaker inductive biases degrade more substantially. Scaling data and model capacity improves interpolation but not necessarily extrapolation, revealing a trade-off between capacity and OOD generalization. Overall, this work provides a systematic, device-consistent benchmark for AI-based GS prediction and practical guidance for selecting reliable surrogates for real-time plasma control and fusion applications.
Guoyang Shi, Zitong Zhang, Siqi Ding +5
Beijing ENN Fusion Energy Science and Technology Co., Ltd., Beijing 101111, China · Beijing Key Laboratory of High Magnetic Field Spherical Torus Fusion Energy, Beijing 101111, China · Hebei Key Laboratory of Compact Fusion, Langfang 065000, China +3
Fast surrogates for tokamak boundary-plasma simulation are typically deterministic regressors mapping a global operating point to a flattened vector of cell values. Near the divertor detachment transition the steady state is not reliably single-valued. A point estimate must average over qualitatively different plasma states, and it arrives with no statement of confidence. Moreover, the flattened vector representation discards the geometric structure of the SOLPS-ITER mesh. This work addresses both problems. We unroll the curvilinear mesh into three fixed-size image tensors whose layout preserves cell adjacency and inverts exactly, letting a convolutional network act on the geometry without loss of information. A conditional flow matching model, well suited to highly sensitive systems, is then trained on this representation. The result is an efficient, scalable surrogate that captures multiple plausible outcomes even at sensitive operating points. Along a gas-puff scan, the predictive distribution splits into a hot and a cold mode across an early regime transition. A further check on synthetic data with an injected bifurcation of known size confirms the model recovers both branches rather than their average.
Gabriele Gianuzzo, Stefan Dasbach, Fleur Hendriks +2