Partial State-Feedback Reduced-Order Switching Predictive Models for Next-Generation Optical Lithography Systems

Basic Information

• Paper ID: 2402.13928
• Title: Partial State-Feedback Reduced-Order Switching Predictive Models for Next-Generation Optical Lithography Systems
• Authors: Raaja Ganapathy Subramanian (Eindhoven University of Technology & ASML NV), Barry Moest (ASML NV), Bart Paarhuis (ASML NV)
• Classification: math.OC (Mathematical Optimization and Control)
• Submission Date: Preprint submitted to Control Engineering Practice, October 15, 2025
• Paper Link: https://arxiv.org/abs/2402.13928

Abstract

This paper proposes a partial state-feedback reduced-order switching predictive model to support next-generation optical lithography technology roadmaps. The approach addresses the trade-off between increasing measurement quantities to improve overlay accuracy and the associated challenges, including elevated measurement noise, reduced throughput, and overlay/placement errors under uncertain operating conditions. By minimizing chip placement errors and reducing unnecessary measurements, the method enhances system performance and throughput. The solution employs a streamlined model with adaptive switching logic to manage time-varying uncertainties caused by fluctuating operating conditions. The approach is implemented on state-of-the-art lithography scanners to mitigate the spatiotemporal dynamics of reticle heating, serving as a representative industrial application. Reticle heating deteriorates with increased throughput, introducing spatiotemporal distortions that directly reduce chip placement accuracy. Experimental results demonstrate significant improvements: placement error reduction of 2-3 times and throughput improvement of 0.3 seconds per wafer.

Research Background and Motivation

Problem Background

1. Importance of Lithography Technology: Lithography is the most critical technology in semiconductor manufacturing processes, determining the manufacturing scale of chip components and directly affecting integrated circuit production quality.
2. Technical Challenges:
   • Minimum feature sizes ranging from 3nm to 500nm, requiring sub-nanometer precision
   • Contradiction between high throughput demands and high precision requirements
   • Impact of spatiotemporal dynamic behavior on system performance
3. Limitations of Existing Methods:
   • Simulation assumptions do not match real-time system behavior
   • Sensitivity to uncertainties and external disturbances
   • Difficulty ensuring global closed-loop stability in real-time operations
   • Trade-off between physical modeling accuracy and model simplicity
   • Constraints on modularity, scalability, and implementation costs

Research Motivation

This paper aims to bridge the gap between theoretical control concepts and practical implementation in lithography systems, proposing a partial state-feedback, reduced-order switching predictive control method while maintaining throughput and overlay performance.

Core Contributions

1. Reduced-Order Model Design: Proposes using reduced-order linear time-invariant models derived from nominal large-scale system dynamics (finite element models), ensuring computational efficiency without compromising model fidelity through selective decoupling of critical spatiotemporal dynamics.
2. Adaptive Switching Mechanism: Introduces a switching logic mechanism to adapt to uncertainties caused by physical effects, improving robustness.
3. Stability Guarantees: Proves that the proposed control strategy ensures global uniformly ultimately bounded asymptotic stability (GUAS), supported by the small-gain theorem for nonlinear systems.
4. Industrial Verification: Implemented and validated on state-of-the-art lithography systems, focusing on mitigating spatiotemporal distortions caused by reticle heating.

Methodology Details

Task Definition

Inputs:

• System state $x \in \mathbb{R}^n$ (predicted complete reticle deformation)
• Exogenous input $u_e \in \mathbb{R}^m$ (heat sources)
• Sparse sensor data $y \in \mathbb{R}^q$

Outputs:

• Output $z \in \mathbb{R}^p$ related to placement error

Constraints:

1. Predicted misalignment remains bounded: $\|z\| \leq \varepsilon$ (sub-nanometer level)
2. Reduce measurement time by selectively skipping measurements of certain mark sets

Model Architecture

State-Space Formulation

The system's state-space formulation is expressed as:

$\hat{P} := \begin{cases} \dot{x}_i = A_i x_i + B_i u_e + B_f u_f \\ u_f = \Gamma(.)y \\ u_i = C_i x_i \end{cases}$

Where:

• $x_i$: internal deformation state
• $u_e$: external heating profile
• $u_f$: feedback signal from alignment sensors
• $\Gamma(.)$: function converting sparse feedback measurements to appropriately dense layout
• $C_i$: output mapping from $x_i$ to $u_i$

Four-Step Design Methodology

Step 1 - Scheduler Integration and Closed-Loop Feedback: The scheduler $\Phi$ is defined as: $\Phi = \{M_i | I \in R_i, \forall i \in \mathbb{Z}\}$

Where $I \in \{y, u_e, u_\Delta\}$ represents the historical information set.

Step 2 - Nominal Reduced-Order Model: Employs Krylov moment-matching method for parameterized model reduction: $M_i := \begin{cases} \dot{x}_i = A_i x_i + B_i u_e \\ u_i = C_i x_i \end{cases}$

Step 3 - Partial Feedback Integration: Integrates partial state feedback into the reduced-order model by explicitly introducing the feedback input channel $B_f$.

Step 4 - Uncertainty Handling: Uses a centralized approach to represent the model as: $\hat{P}(\Phi|I) = M_n + M_\Delta = M_n + \Delta_m(\Phi|I)$

Technical Innovations

1. History-Driven Switching Strategy: Unlike variable-gain control based solely on feedback signal magnitude, this method uses historical data $I$ to guide model selection.
2. Preservation of Physical Interpretability: Maintains the original physical interpretation of the reduced-order model through the Krylov moment-matching method.
3. Lur'e-Type System Framework: Transforms the system into a Lur'e-type system, combining linear dynamic components with nonlinear/uncertain components.

Experimental Setup

Experimental Platform

• Equipment: ASML Twinscan state-of-the-art lithography scanner
• Application Scenario: Reticle heating mitigation
• Test Conditions: 2 batches, 16 wafers per batch

Evaluation Metrics

1. Placement Error: Overlay placement error in x-y directions
2. Throughput Improvement: Time savings per wafer
3. Stability: Performance consistency under different operating conditions

Comparison Methods

1. Traditional Method: Current standard sensor-based method
2. Linear Method: Linear version using only the nominal model $M_n$

Implementation Details

• Uses 3 moments to describe nominal and uncertain reticle heating models
• Skips edge alignment marks, using only top-bottom alignment marks
• Saves approximately 0.3 seconds per wafer per measurement

Experimental Results

Main Results

1. Placement Error Improvement:
   • 2-3 times reduction in x-y direction placement error compared to traditional methods
   • 2 times performance improvement in x-y direction for the first 2 wafers
   • Maintains stable and improved accuracy under both nominal and uncertain conditions
2. Throughput Improvement:
   • Saves 0.3 seconds per wafer
   • Can increase throughput by up to 7 wafers per hour
3. Robustness Verification:
   • Linear method shows 3 times performance degradation under uncertain conditions
   • Proposed method maintains consistent performance across different image regions

Verification Across Different Image Regions

In experiments with smaller exposure areas:

• Proposed method achieves 3 times performance improvement under both nominal and uncertain conditions
• Linear method shows 2-3 times performance degradation
• Validates the robustness and adaptability of the method

Stability Analysis

Global uniformly ultimately bounded asymptotic stability (GUAS) is proven through the small-gain theorem, satisfying:

• Reticle heating physics ensures bounded dynamics
• Both nominal and uncertain models satisfy stability conditions
• Equilibrium point $x^*$ (where $\|z\| = 0$) is GUAS

Related Work

Lithography System Control

• Traditional linear predictive control methods are limited by the "waterbed effect"
• Nonlinear approaches include variable-gain control, filtering, and switching controllers
• Existing methods often neglect spatiotemporal dynamics

Model Reduction Techniques

• Balanced truncation, Krylov subspace methods, proper orthogonal decomposition (POD)
• Existing methods reduce dimensionality but lose physical interpretability
• This paper's method maintains physical interpretation while achieving computational efficiency

Switching System Control

• History-based switching strategies
• Event-driven control methods
• Lur'e system stability theory

Conclusions and Discussion

Main Conclusions

1. Successfully developed partial state-feedback reduced-order switching predictive models
2. Achieved simultaneous improvement in accuracy and throughput
3. Verified robustness under uncertain operating conditions
4. Provided theoretical stability guarantees

Limitations

1. Method applicability depends on specific physical system characteristics
2. Switching logic design requires domain expertise
3. Model reduction may affect accuracy under certain extreme conditions

Future Directions

1. Unified Predictive Framework: Extend to spatial and temporal variations across the entire system
2. Intelligent Measurement Optimization: Integrate prediction, learning, and intelligence to reduce unnecessary measurements
3. Practical Assumption Validation: Narrow the gap between theoretical concepts and practical applications

In-Depth Evaluation

Strengths

1. Strong Industrial Relevance: Directly addresses practical problems in semiconductor manufacturing
2. Solid Theoretical Foundation: Provides rigorous stability analysis
3. Comprehensive Experimental Validation: Verified on real industrial systems
4. Methodological Innovation: Novel approach combining reduced-order modeling, switching control, and partial feedback
5. High Practical Value: Simultaneously improves accuracy and throughput with direct economic value

Weaknesses

1. Limited Applicability Scope: Primarily addresses reticle heating issues; generalization to other applications requires verification
2. Complexity Management: Design and tuning of switching logic may require substantial expertise
3. Limited Comparative Experiments: Lacks in-depth comparison with other advanced control methods
4. Long-Term Stability: Insufficient discussion of performance degradation during extended operation

Impact

1. Academic Value: Provides new theoretical framework for spatiotemporal system control
2. Industrial Value: Direct application to multi-billion-dollar semiconductor equipment
3. Technology Dissemination: Method extensible to other precision manufacturing systems
4. Reproducibility: Provides clear implementation steps and theoretical foundation

Applicable Scenarios

1. Precision Manufacturing Systems: Manufacturing processes requiring sub-nanometer precision
2. Thermal Management Applications: Systems involving thermal deformation control
3. High-Throughput Requirements: Scenarios requiring efficiency improvement while maintaining precision
4. Uncertain Environments: Industrial systems with frequently changing operating conditions

References

The paper cites 45 relevant references covering multiple fields including control theory, lithography technology, model reduction, and system identification, providing a solid theoretical foundation and technical background for the research.

Overall Assessment: This is a high-quality engineering application paper that successfully applies advanced control theory to practical industrial problems, achieving a good balance between theoretical rigor and practical value. The contribution lies not only in technical innovation but also in providing practical solutions for the critical semiconductor manufacturing industry.