Virtual Metrology for Plasma Etching: A Two-Stage Framework for Sensor Data Compression and Prediction

Technical Summary: Virtual Metrology for Plasma Etching: A Two-Stage Framework for Sensor Data Compression and Prediction

Problem Statement
Plasma etching is a critical yet difficult-to-monitor process in semiconductor manufacturing, governed by nonlinear plasma chemistry, sheath dynamics, and surface reactions. While physical metrology (e.g., SEM, ellipsometry) provides accurate measurements of key quality attributes like critical dimension (CD), etch depth (ED), and profile shape, it is inherently sparse, delayed, and costly. Typically, only a small fraction of wafers in a lot are measured, creating an asymmetry where process control requires dense wafer-level information but physical metrology provides only delayed, sparse labels. Conversely, plasma etching tools generate abundant, high-dimensional, time-resolved in situ data (e.g., Optical Emission Spectroscopy [OES], RF/electrical signals, Equipment Engineering System [EES] variables). Virtual Metrology (VM) aims to bridge this gap by estimating wafer quality from these sensor data. However, the field faces significant challenges: the input data are noisy, redundant, and highly drift-prone, and existing literature often treats preprocessing and modeling as monolithic blocks, obscuring the specific architectural requirements for compressing complex plasma signals into robust predictive representations.

Methodology
This paper presents a systematic review of 105 plasma-etch VM studies published between 2012 and 2025. The authors organize the field around a two-stage framework (illustrated in Figure 2) to explicitly separate data compression from prediction:

• Stage 1: Data Compression and Representation: This stage transforms raw, high-dimensional sensor data into a compact, informative representation. The review categorizes methods into:
  • Feature Selection: Filter (statistical scoring), Wrapper (model-based search), and Embedded (regularization-based) approaches.
  • Feature Transformation: Statistical summaries, linear projections (PCA, PLS), spectral/time-frequency transforms (FFT, Wavelets), clustering, and deep representation learning (CNNs, Autoencoders).
  • Hybrid Strategies: Sequential compositions or combined optimizations that handle multi-modal data (OES, RF, EES) differently before merging them.
• Stage 2: Prediction and Decision Making: This stage maps the Stage 1 representation to target variables. The review covers:
  • Regression: Linear/PLS models, regularized methods (LASSO, Elastic Net), kernel methods (SVR, GPR), and neural networks (MLP, LSTM).
  • Classification/Detection: Endpoint detection (EPD), fault detection (FDC), and profile classification using SVMs, tree ensembles, and deep learning (Capsule Networks, Transformers).
  • Uncertainty & Interpretability: Probabilistic models (GPR, RVM) for confidence estimation and methods for physical interpretability (knowledge-guided feature selection, Plasma-Information variables).

The review also analyzes system-level couplings, including multi-task learning (shared representations for multiple targets), cross-process dependencies, and chamber-to-chamber transferability.

Key Contributions

1. Structural Framework: The primary contribution is the explicit articulation of a two-stage pipeline for plasma-etch VM. This distinguishes the construction of compact, physically meaningful representations (Stage 1) from the final prediction task (Stage 2), clarifying that the representation step is often the bottleneck for robustness and deployability.
2. Systematic Taxonomy: The paper provides a comprehensive taxonomy of 105 studies, cross-tabulating compression strategies with prediction models (Table 3). It reveals that while Stage 2 predictors have modernized (increasing use of NNs), Stage 1 representations remain dominated by classical methods (feature selection/transformation), with deep representation learning still in the minority.
3. Identification of Gaps: The authors identify four critical gaps hindering industrial adoption:
   • Data Gap: The imbalance between high-dimensional sensors and sparse, delayed labels.
   • Architecture Gap: The lack of clear separation between compression and prediction, and the underutilization of multi-task/shared-representation architectures.
   • Deployment Gap: Insufficient attention to model drift, tool mismatch, latency, and uncertainty quantification in production environments.
   • Benchmarking Gap: The absence of common datasets and standardized reporting protocols, making cross-study comparison difficult.
4. Open Dataset: The authors released a structured dataset of the 105 reviewed studies as an open resource to facilitate future benchmarking.

Results and Findings

• Temporal Trends: Research activity has grown from 2012 to 2025, shifting from classical statistical methods (2012–2015) to machine learning (2016–2019) and deep learning/system integration (2020–present).
• Geographical Distribution: Publication output is led by South Korea (45 papers) and the United States (17 papers). The remainder is more evenly distributed, with Ireland and Italy contributing eight papers each, Japan seven, and Taiwan and China five each. Collectively, European groups (Ireland, Italy, Germany, and Belgium) account for about one-fifth of the corpus (22 papers). The review notes a partial mismatch between publication volume and manufacturing capacity, suggesting that much industrial VM work in high-capacity regions like Taiwan and China remains proprietary and undisclosed.
• Methodological Coupling: The review finds that deep learning representations are almost exclusively paired with neural network predictors, whereas classical compression methods are paired with linear or kernel-based models. This suggests a lack of cross-pollination between advanced representation learning and simpler, more interpretable predictors.
• Physical Meaning: The most robust approaches often incorporate "Plasma-Information" (PI) variables or knowledge-guided feature selection, linking sensor data to physical mechanisms (e.g., radical density, sheath voltage) rather than relying solely on abstract statistical components.

Significance and Claims
The paper claims that the main challenge in plasma-etch VM is not merely achieving high prediction accuracy, but constructing compact, robust, and physically meaningful representations from noisy, drift-prone data. The authors argue that a mature VM system must treat representation, prediction, and deployment constraints as jointly coupled design choices.

The significance of this work lies in its shift from viewing VM as a simple regression problem to viewing it as a system-level engineering challenge. By separating compression from prediction, the framework highlights that:

• Interpretability is crucial for engineer trust and is best achieved through knowledge-guided compression rather than post-hoc explanation of black-box models.
• Uncertainty quantification is essential for deciding when to defer to physical metrology, particularly given the sparse label problem.
• Adaptability (via transfer learning, domain adaptation, and just-in-time modeling) is required to handle chamber drift and tool-to-tool variation.

The authors modestly conclude that while VM has demonstrated potential for endpoint detection, fault classification, and quality estimation, wider industrial deployment depends on addressing the identified gaps in label efficiency, architectural clarity, and robustness to distribution shifts. They emphasize that future progress requires VM systems that are not just accurate predictors, but adaptive, uncertainty-aware components integrated into Advanced Process Control (APC) and Fault Detection and Classification (FDC) workflows.