AI4Plasma I: Data-Efficient Surrogate Modeling of Capacitively Coupled Plasma Fields

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

TL;DR

Simulating plasma fields for semiconductor manufacturing is computationally grueling, often taking hours or days per case. This paper systematically evaluates five neural network architectures to find the most "data-efficient" surrogate model—the one that achieves high accuracy with the fewest expensive reference simulations. The Fourier Neural Operator (FNO) emerges as the top all-rounder, particularly in low-data regimes, while U-Net and Transolver offer specialized advantages for high-accuracy interior and boundary predictions, respectively.

What Problem Does This Paper Address?

In the race to shrink semiconductor features, engineers rely on Capacitively Coupled Plasma (CCP) simulations to optimize etching and deposition processes. However, these physics-based models are a bottleneck: a single 2D simulation in this study requires an average of 3.8 CPU core-hours.

Surrogate models (AI "stand-ins") can provide near-instant predictions, but they require a training dataset of these expensive simulations. The authors tackle a critical practical question: Which AI architecture reaches a target accuracy (e.g., 1% error) with the minimum number of training samples? This "data efficiency" is the key to making AI-driven plasma design feasible in the real world.

Methods

The researchers compared five representative architectures that encode spatial information in fundamentally different ways:

  1. Coordinate MLP: A pointwise approach where a simple network predicts field values based on coordinates and operating parameters.
  2. PCA-Net: A global low-rank approach that predicts coefficients for a pre-calculated Principal Component Analysis (PCA) basis.
  3. U-Net: A local multiscale approach using convolutions and skip connections to capture features at different resolutions.
  4. Fourier Neural Operator (FNO): A global spectral approach that learns to manipulate data in the Fourier domain.
  5. Transolver: An adaptive nonlocal approach that uses attention-based "slicing" to group related spatial points regardless of their physical distance.
Figure 2: Spatial representations of the five surrogate architectures. Coordinate MLP predicts fields pointwise, PCA-Net reconstructs fields using a global PCA basis, and U-Net, FNO, and Transolver exchange spatial information through multiscale convolutions, Fourier modes, and adaptive spatial grouping, respectively.

Datasets

The models were tested on two standard benchmarks for Argon discharges:

  • 1D Planar CCP: Resolves axial bulk-sheath dynamics (128 cells).
  • 2D Axisymmetric GEC CCP: A more complex reactor geometry resolving radial nonuniformity and geometric transitions (4,624 active cells).

The datasets were generated by varying three key operating parameters: RF voltage ($U_{RF}$), gas pressure ($P$), and driving frequency ($f$).

Figure 1: Geometries and principal boundary conditions of the reference CCP benchmarks: (a) 1D planar CCP and (b) 2D axisymmetric GEC CCP.

Key Insights and Ideas

  • Inductive Biases Matter: FNO’s "spectral bias" makes it exceptionally good at capturing the relatively smooth but complex global structures of plasma fields with very little data. In 1D, it reached the 1% error target with just 50 cases, while U-Net needed 200.
  • The "Boundary" Challenge: Predicting plasma behavior at the edges of the operating window (e.g., very high voltage or low pressure) is much harder than predicting the "interior." While FNO dominates the interior, the Transolver proved more robust at the parameter boundaries when high accuracy (1% error) was required.
  • PCA-Net’s Hidden Gap: The authors found that PCA-Net's error isn't just caused by truncating the basis (low-rank representation). Instead, the primary difficulty is accurately mapping the operating parameters to the correct PCA coefficients.
  • Structural Importance: Ablation studies showed that skip connections are vital for U-Net to resolve sharp gradients near sheaths, and adaptive slicing allows Transolver to capture nonlocal interactions across the reactor chamber.

Key Results

  • Overall Winner: FNO required the lowest or joint-lowest number of reference cases in 7 out of 8 primary comparisons.
  • CPU Savings: For a 2% error target in 2D boundary cases, using FNO instead of U-Net saves approximately 760 CPU core-hours of data generation.
  • High-Accuracy Niches:
    • U-Net achieved the lowest absolute error (0.126%) for 2D interior fields when given the full 800-case budget.
  • Transolver was the most efficient for the strict 1% 2D boundary target, requiring 400 cases compared to FNO's 800.
Table 1: Reference-simulation setup, data sizes and computational cost for the two CCP bench- marks.

Limitations and Future Work

The authors note that this study focused on fixed geometries and a relatively simple Argon chemistry. Real-world industrial applications often involve complex gas mixtures (like $CF_4$ or $O_2$) and varying reactor designs.

Future work aims to extend this data-efficiency framework to:

  • More complex chemical reaction sets.
  • Varying reactor geometries (moving beyond fixed grids).
  • Wafer-specific metrics, such as ion energy distributions, which are critical for predicting actual etch profiles.

Summary updated

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.
BibTeX
@misc{zou2026ai,
  author = {Zhenyu Zou and Junjue Liao and Hao Yang and Luowei Yin and Chenyu Wang and Yufeng Wei and Zhichao Peng and Chun Chen and Daniel Zhengyu Huang and Kangxin Liu and Haoran Guan and Rubing Han and Yipei Chen and Peiting You and Yuwei Fan},
  title = {{AI4Plasma I: Data-Efficient Surrogate Modeling of Capacitively Coupled Plasma Fields}},
  year = {2026},
  eprint = {2611.15002},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.15002},
}
Figure