Physics-Informed Neural Systems for the Simulation of EUV Electromagnetic Wave Diffraction from a Lithography Mask

This paper introduces a novel hybrid Waveguide Neural Operator (WGNO) and evaluates Physics-Informed Neural Networks (PINNs) for simulating Extreme Ultraviolet (EUV) wave diffraction from lithography masks, demonstrating that these neural systems achieve state-of-the-art accuracy with significantly faster inference times and strong generalization compared to traditional numerical solvers.

The Gist

The Big Picture: The "Impossible" Puzzle of Chip Making

Imagine you are trying to print a microscopic blueprint onto a silicon wafer to make a computer chip. This is done using Extreme Ultraviolet (EUV) lithography, which is like using a super-powerful, invisible flashlight to burn patterns onto the chip.

However, there's a catch. The "mask" (the stencil that holds the pattern) is so tiny and complex that the light doesn't just travel in straight lines. It bends, bounces, and interferes with itself (a phenomenon called diffraction). It's like trying to shine a laser through a complex maze of mirrors; the light gets messy, and the pattern on the chip doesn't look exactly like the pattern on the mask.

To fix this, engineers use a process called Optical Proximity Correction (OPC). They have to tweak the mask design so that, after the light gets messy, the final result on the chip is perfect.

The Problem: To do this tweaking, engineers need to simulate exactly how the light behaves. The current way to do this is like trying to solve a massive, 3D jigsaw puzzle where every piece is a tiny wave of light. It requires supercomputers and takes hours or even days to get one answer. This slows down the creation of new, faster chips.

The Solution: Teaching Light to "Guess" with Physics

The authors of this paper propose a new way to solve this puzzle using Artificial Intelligence (AI), specifically a type called Physics-Informed Neural Networks (PINNs) and a new invention they call the Waveguide Neural Operator (WGNO).

Here is how they work, using analogies:

1. The Old Way: The "Brute Force" Calculator

Imagine you are trying to predict the weather. The old method (like the standard Waveguide Method used in the industry) is like measuring the temperature, humidity, and wind speed at every single inch of the planet, then doing a complex math calculation for every single point to see how the air moves. It's incredibly accurate, but it's slow and exhausting.

2. The First AI Attempt: The "Smart Student" (PINN)

The first AI approach they tried is like a very smart student who has been given the rules of physics (the laws of how light moves) but no textbook answers.

• How it works: The student is told, "You must follow these rules of light." The AI tries to guess the answer, checks if it broke the rules, and corrects itself.
• The Result: It's faster than the supercomputer, but the student sometimes gets confused by the complex 3D shapes of the mask. It's like a student who understands the theory but struggles with the specific, messy details of the exam. It's okay for simple problems, but not perfect for the real-world chip masks.

3. The New Champion: The "Master Chef" (WGNO)

This is the paper's main breakthrough. The authors realized that the "brute force" method has a specific structure (it breaks the problem down into layers, like a sandwich). They decided to build an AI that learns the chef's recipe, not just the ingredients.

• The Analogy: Imagine the "brute force" method is a chef who measures every single grain of salt and sugar for every single cake.
• The WGNO is a master chef who has learned the pattern of the recipe. Instead of measuring every grain, the chef knows exactly how the layers interact.
• How it works: The AI is trained to look at the "ingredients" (the mask design and the light wavelength) and instantly predict the "outcome" (how the light will diffract) by skipping the slow, heavy math calculations that usually take the most time.
• The Magic: It doesn't just memorize specific masks; it learns the physics of the layers. This means if you show it a mask it has never seen before, it can still predict the result with amazing accuracy because it understands the underlying "recipe" of light.

The Results: Speed vs. Accuracy

The researchers tested their new "Master Chef" (WGNO) against the old "Brute Force" method and the "Smart Student" (PINN) on two types of light wavelengths (13.5 nm and 11.2 nm).

• Accuracy: The WGNO was almost as perfect as the slow supercomputer method. The "Smart Student" (PINN) was good, but made noticeable errors on the complex 3D masks.
• Speed: This is where the WGNO shines.
  • The old method takes seconds to minutes to solve a problem.
  • The WGNO takes milliseconds.
  • It is 200 times faster than the rigorous method.

Why Does This Matter?

Think of designing a new chip as trying to bake a cake for a million people, but you only have one oven and you have to get the recipe perfect before you start baking.

• Before: You had to run a slow, expensive simulation for every tiny change in the recipe. It took forever to get the cake right.
• Now: With the WGNO, you can simulate thousands of recipe changes in the time it used to take to do one. You can instantly see how a tiny tweak to the mask will affect the final chip.

The Takeaway

The authors have created a new AI tool that acts like a super-fast, physics-savvy shortcut. It doesn't just guess; it understands the laws of light so well that it can predict how light will bend around complex 3D structures in a fraction of a second.

This means chip manufacturers can design better, faster, and smaller computer chips much more quickly, helping to keep Moore's Law (the idea that chips get faster every two years) alive for another generation. It's a "state-of-the-art" tool that turns a days-long calculation into a blink-of-an-eye prediction.

Technical Summary

1. Problem Statement

The paper addresses the computational challenge of simulating Extreme Ultraviolet (EUV) electromagnetic wave diffraction from 3D lithography masks.

• Context: In semiconductor manufacturing, EUV lithography (using wavelengths like 13.5 nm and promising 11.2 nm) is critical for creating smaller transistors. However, due to diffraction and interference, the pattern on the mask does not perfectly match the pattern on the wafer.
• The Bottleneck: Accurate simulation requires solving Maxwell's equations for complex 3D mask structures (multi-layered absorbers and mirrors). Traditional rigorous solvers (like the Waveguide method or Finite Element Method) are highly accurate but computationally expensive, requiring massive resources and time. This slows down the Optical Proximity Correction (OPC) workflow, which is essential for mask design.
• Goal: To develop deep learning-based surrogates that maintain high accuracy while drastically reducing inference and training times, enabling faster mask optimization.

2. Methodology

The authors propose and compare three distinct approaches to solving the diffraction problem:

A. Baseline Numerical Solvers

• Finite Element Method (FEM): Implemented using the open-source solver FreeFem++. It discretizes the domain into a mesh. While robust, it is computationally heavy for fine wavelength features.
• Waveguide (WG) Method: A high-fidelity reference solver. It exploits the layered structure of the mask, representing the field as a superposition of waveguide modes within each layer. The most computationally expensive step is solving a large linear algebraic eigenvalue problem to determine mode amplitudes at layer interfaces.

B. Physics-Informed Neural Networks (PINNs)

• Approach: An unsupervised learning method where a Multi-Layer Perceptron (MLP) approximates the electromagnetic field components ($H_x, H_y$).
• Training: The network is trained by minimizing a composite loss function consisting of:
  1. Residual Loss ($L_r$): The error in satisfying the governing partial differential equations (Maxwell's equations) at random collocation points.
  2. Boundary Loss ($L_{bc}$): The error in satisfying periodicity and Sommerfeld radiation conditions.
• Features: Uses Fourier feature embeddings for input coordinates to strictly enforce periodic boundary conditions.

C. Waveguide Neural Operator (WGNO) – The Novel Contribution

• Concept: A hybrid architecture that replaces the most computationally intensive part of the traditional Waveguide method with a neural network.
• Mechanism:
  • The WG method involves solving a linear system $\hat{M}A = R$ to find mode amplitudes ($A$) given the mask parameters (permittivity Fourier coefficients) and incident field ($R$).
  • The WGNO trains an MLP to learn the direct mapping from the input parameters ($R$ and mask permittivity coefficients) to the solution vector ($A$).
  • Once the neural network predicts $A$, the rest of the WG method (field reconstruction via Fourier series) is performed analytically.
• Advantage: It inherits the physical structure of the WG method (ensuring physical consistency) but bypasses the expensive matrix inversion/eigenvalue solving step during inference.

3. Key Contributions

1. Introduction of WGNO: A novel hybrid operator that integrates deep learning into the rigorous Waveguide method, specifically targeting the computational bottleneck of the linear system solution.
2. Unsupervised Training: Demonstrated that both PINNs and WGNOs can be trained without labeled datasets (supervised data), relying solely on the governing physical equations.
3. Generalization Capability: Proved that the neural operators, particularly WGNO, possess strong generalization properties, delivering high accuracy for problem parameters (e.g., hole sizes, layer thicknesses) not seen during training.
4. Multi-Wavelength Validation: Validated the methods on both current industrial standards (13.5 nm) and next-generation wavelengths (11.2 nm).

4. Experimental Results

The methods were evaluated on 2D test cases with analytical solutions, realistic 2D masks, and full 3D masks.

• Accuracy (Relative $L_2$ Error):
  • PINN: Achieved reasonable accuracy for simple cases but struggled with complex 3D structures and dielectric layers, showing errors in the range of $10^{-2}$ to $10^{-1}$ (1%–10%).
  • WGNO: Achieved state-of-the-art accuracy, with errors as low as $10^{-7}$ to $10^{-8}$, comparable to the rigorous WG solver (machine precision).
• Speed and Efficiency:
  • Training Time: WGNO trained significantly faster than PINN (e.g., ~4 seconds vs. ~1000 seconds for 2D test problems).
  • Inference Time: WGNO was orders of magnitude faster than FEM and significantly faster than PINN.
    • 3D Mask Inference: WGNO achieved inference times of $\approx 2 \times 10^{-4}$ seconds.
    • Speedup: The WGNO provided a >200x speedup compared to the rigorous WG solver while maintaining high fidelity.
• Generalization:
  • For unseen mask parameters (hole widths), the WGNO maintained a generalization measure ($\mu$) close to 1.0 (e.g., 0.999 for 100 partitions), indicating it predicts unseen scenarios with accuracy nearly identical to trained scenarios.

5. Significance and Conclusion

• Accelerating Lithography Design: The WGNO offers a highly efficient solution for the Optical Proximity Correction (OPC) workflow. By reducing simulation time from seconds/minutes to microseconds, it enables rapid iteration in mask design, which is critical for Moore's Law scaling.
• Scalability: The method scales effectively from 2D to complex 3D problems without a proportional increase in computational cost.
• Broader Impact: While focused on EUV lithography, the authors argue that the WGNO architecture is applicable to other complex electromagnetic structures, such as metamaterials and composite materials, where rigorous solvers are currently too slow for design optimization.

In summary, the paper demonstrates that hybrid physics-informed neural operators (specifically WGNO) outperform both pure data-driven neural networks (PINNs) and traditional numerical solvers, offering the best balance of rigorous accuracy and computational speed for next-generation semiconductor manufacturing.