AI4Plasma II: Geometry-Aware Surrogate Modeling for Variable Electrode Gap GEC Plasmas

Hao Yang, Zhenyu Zou, Wenshuai Jie, Luowei Yin, Yufeng Wei, Daniel Zhengyu Huang, Zhichao Peng, Chun Chen, Kangxin Liu, Rubing Han, Haoran Guan, Yipei Chen, Peiting You, Yuwei FanarXiv 2026 · arXiv

TL;DR

Researchers have developed a way to rapidly predict plasma behavior in semiconductor manufacturing tools even when the physical dimensions of the tool change. By comparing six "geometry-aware" neural network architectures and introducing a specialized physics-weighted loss function, they achieved high-accuracy predictions with 10x less training data than standard methods.

What Problem Does This Paper Address?

In semiconductor fabrication, capacitively coupled plasmas (CCPs) are the workhorses of etching. To get perfect results, engineers must fine-tune operating parameters like pressure and voltage, but also geometric ones—specifically the electrode gap.

The challenge is that high-fidelity fluid simulations are computationally expensive, taking over 5 hours per case. While machine learning "surrogate models" can provide near-instant predictions, they usually break when the geometry changes. If you change the electrode gap, the computational grid shifts and the domain size alters, making it impossible for standard neural networks to "align" the data across different samples. This paper tackles how to build models that are truly geometry-aware and data-efficient.

Figure 1: Two-dimensional axisymmetric computational domain and mesh of the variable-gap GEC configu- ration. The electrodes are separated from the surrounding grounded wall by thin annular insulators.

Methods

The authors systematically compared six different strategies for handling variable geometry:

  1. Geo-FNO-A (Prescribed Mapping): Uses a mathematical formula to "stretch" every geometry into a standard reference box before processing it with a Fourier Neural Operator (FNO).
  2. Geo-FNO-L (Learned Mapping): Similar to above, but the network learns its own internal coordinate transformation.
  3. eDAFNO (Extended Domain): Places all geometries into one large fixed box and uses a "mask" to tell the network where the plasma is.
  4. C-MLP (Continuous Representation): Treats the plasma field as a continuous function of $(r, z)$ coordinates.
  5. PCNO (Point Cloud): Processes the simulation mesh as an irregular set of points.
  6. Transolver (Physics-Attention): A Transformer-based architecture that groups grid cells into "physics slices" to learn global interactions.

A key innovation is the Charge-Separation-Weighted Loss. Because the most critical plasma physics happens in the thin "sheath" layers near the wafer (where ions and electrons separate), the authors modified the training loss to penalize errors in these regions more heavily.

Datasets

The study uses a dataset generated from 2D axisymmetric fluid simulations of a Gaseous Electronics Conference (GEC) reference cell.

  • Total Cases: 1,200 (1,000 training, 100 validation, 100 test).
  • Parameters: Four variables were sampled using Latin Hypercube Sampling: gas pressure, RF voltage, driving frequency, and the electrode separation gap.
Table 1: Operating-parameter ranges used for dataset generation.

Key Insights and Ideas

  • Whole-field error is a trap: A model might have a 1% error across the entire chamber but fail to capture the ion density near the wafer. The paper shows that predicting near-wafer profiles is significantly harder than predicting the bulk plasma.
  • Prior knowledge beats raw power: When training data is scarce (e.g., only 100 cases), Geo-FNO-A—which uses a simple prescribed geometric mapping—outperforms more complex models that try to learn the geometry from scratch.
  • The "Sheath" is the secret: By forcing the model to care about charge separation (the physics that defines the sheath), they achieved a massive boost in accuracy where it matters most: at the wafer surface.
Figure 2: Ion-density distributions at electrode gaps of (a) 2, (b) 4, and (c) 6 cm, with p = 0.316 Torr, VRF = 200 V, and fRF = 40.68 MHz. All panels share the same color scale.

Key Results

  • Data Efficiency: To reach a 2% mean error across all plasma fields, Geo-FNO-A and Transolver required only 100 training cases. Other models needed up to 1,000.
  • The 10x Improvement: For the difficult task of predicting ion density profiles near the wafer with 5% error, the new charge-separation weighting reduced the required training data for Geo-FNO-A from 1,000 cases down to just 100.
  • Model Recommendations:
    • Use Geo-FNO-A if you have a clear geometric mapping and limited data.
  • Use Transolver if you have abundant data and need the absolute highest precision.

Limitations and Future Work

  • Simple Chemistry: The study focused on pure Argon plasma. Real-world industrial processes use complex gas mixtures (like $CF_4/O_2$) with dozens of species and chemical reactions.
  • Geometric Complexity: While the electrode gap was varied, the overall "shape" of the chamber remained a simple cylinder. Future work needs to test these models on more complex, non-convex reactor geometries.
  • Steady-State Only: The models currently predict cycle-averaged steady-state fields; capturing the full time-dependent transient behavior remains a challenge.

Summary updated

Abstract

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.
BibTeX
@misc{yang2026ai,
  author = {Hao Yang and Zhenyu Zou and Wenshuai Jie and Luowei Yin and Yufeng Wei and Daniel Zhengyu Huang and Zhichao Peng and Chun Chen and Kangxin Liu and Rubing Han and Haoran Guan and Yipei Chen and Peiting You and Yuwei Fan},
  title = {{AI4Plasma II: Geometry-Aware Surrogate Modeling for Variable Electrode Gap GEC Plasmas}},
  year = {2026},
  eprint = {2611.15584},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.15584},
}
Figure