🔬 optics

Gradient-based inverse lithography for EUV masks via the waveguide method and a physics-informed neural operator

This paper presents a gradient-based inverse lithography framework for EUV masks that integrates differentiable waveguide methods and waveguide neural operators as end-to-end physics engines to automatically recover optimal absorber permittivity, successfully generating mask structures that achieve desired wafer fields for realistic 2D and 3D materials at 11.2 nm.

Original authors: Vasiliy A. Es'kin, Egor V. Ivanov

Published 2026-06-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Vasiliy A. Es'kin, Egor V. Ivanov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to bake the perfect cake, but you can't see the oven. All you can see is the final result on the plate. Your goal is to figure out exactly how much flour, sugar, and eggs you need to mix together to get that perfect cake. This is essentially what the scientists in this paper are doing, but instead of baking, they are designing microscopic patterns for computer chips.

Here is a simple breakdown of their work:

The Big Problem: The "Shadow" Effect

Modern computer chips are so tiny that they are measured in atoms. To print these tiny circuits, manufacturers use a special light called Extreme Ultraviolet (EUV). They shine this light through a "mask" (a stencil) to project the circuit pattern onto a silicon wafer.

However, because the light is so small and hits the mask at an angle, it behaves like water waves crashing against a rocky shore. The 3D shape of the mask creates weird "shadows" and distortions (called diffraction) that ruin the pattern on the chip. It's like trying to draw a straight line with a flashlight that keeps flickering and bending the beam.

The Old Way: Guessing and Checking

To fix this, engineers use a technique called "Inverse Lithography Technology" (ILT). They want to work backward: "What shape should the mask be so that the light lands perfectly on the chip?"

Traditionally, this is like trying to solve a maze by running through it blindfolded. You guess a shape, simulate the light, see if it's wrong, change the shape, and try again. The problem is that simulating the light is incredibly slow and computationally expensive. It's like trying to calculate the weather for the whole planet every time you want to know if you need an umbrella.

The New Solution: A "Smart" Physics Engine

The authors of this paper created a new, faster way to solve this puzzle. They combined two clever ideas:

The "Differentiable" Waveguide: Think of their simulation tool as a super-smart calculator that doesn't just give you an answer; it also tells you how to change the answer to get closer to the goal. Instead of guessing blindly, the system knows exactly which tiny tweak to make to the mask to improve the light pattern. It's like having a GPS that doesn't just tell you "you're lost," but says, "turn left 5 degrees to get back on track."
The "Neural Operator" (The AI Assistant): They also trained a small AI (a neural network) to act as a shortcut. Usually, the hardest part of the calculation is solving a massive system of equations. The AI learns to predict the answer to these equations almost instantly, acting like a seasoned chef who knows exactly how long to bake a cake without needing to check the oven every minute.

How They Tested It

The team tested their method on a virtual 2D and 3D world using a specific light wavelength (11.2 nm, which is even smaller than the standard 13.5 nm used today). They tried three different "ingredients" for the mask's absorbing layer:

TaBN: A common material.
Lanthanum (La): A rare earth metal.
Uranium (U): A heavy metal.

They asked the system to create specific light patterns (like a single bright spot or three bright spots) on the "wafer."

The Results

It Works: The system successfully designed masks that created the exact light patterns they wanted, even with the tricky 3D distortions.
Speed: They tried two ways to describe the mask shape. One was like drawing with a pixel-by-pixel grid (very detailed but slow). The other was like drawing with smooth, mathematical curves (Fourier series). The smooth curve method was 1.31 times faster and produced cleaner, smoother edges that would be easier to manufacture.
Material Matters: They found that Lanthanum created the brightest central spot, while Uranium created a pattern that most closely matched the ideal target shape.
3D Success: They proved this works not just for flat lines, but for complex 3D shapes, which is crucial for real-world chips.

The Bottom Line

This paper presents a new "smart" way to design the stencils (masks) for the world's most advanced computer chips. By using a mix of rigorous physics and smart AI shortcuts, they can figure out the perfect mask shape much faster and more accurately than before. This ensures that the tiny circuits on our future computers are printed with perfect clarity, avoiding the "blurry shadows" that usually plague these ultra-small designs.

Technical Summary: Gradient-Based Inverse Lithography for EUV Masks via the Waveguide Method and a Physics-Informed Neural Operator

1. Problem Statement

The paper addresses the Inverse Lithography Technology (ILT) problem for Extreme Ultraviolet (EUV) masks, specifically targeting the manufacturing of advanced semiconductor devices where feature sizes have entered the angstrom regime. Unlike deep ultraviolet systems, EUV lithography (utilizing wavelengths of 13.5 nm and future targets of 11.2 nm) relies on reflective photomasks. The interaction between the mask's three-dimensional topography and obliquely incident electromagnetic waves induces complex diffraction effects that degrade pattern fidelity on the silicon wafer.

While Optical Proximity Correction (OPC) is standard, the most rigorous approach, ILT, formulates mask design as an inverse problem: finding the ideal mask topology (specifically the absorber permittivity distribution) to generate a desired intensity distribution on the wafer. Traditional pixel-based ILT is computationally prohibitive due to the repeated invocation of full-wave electromagnetic simulations. Furthermore, existing deep learning surrogates often require large pre-computed datasets and struggle to generalize across realistic mask geometries.

2. Methodology

The authors propose a gradient-based ILT framework that treats the forward diffraction model as a differentiable physics engine. The core methodology integrates two primary components:

A. Differentiable Waveguide Method

The forward problem is governed by time-harmonic Maxwell equations. The authors employ the waveguide method, which transforms the Helmholtz equation into an algebraic eigenvalue problem for each layer by separating transverse and longitudinal dependencies. This involves:

Expanding the magnetic field component into a series of waveguide modes.
Enforcing continuity of tangential field components at layer interfaces to form a global linear system.
Solving this system to determine the reflected field.

Crucially, the entire forward operator is made differentiable with respect to the design variables (permittivity) via automatic differentiation. This allows for the calculation of gradients without finite-difference approximations, significantly reducing computational cost compared to traditional methods.

B. Neural Operator Integration (WGNO)

To address the computational bottleneck of solving the global linear system, the authors incorporate the Waveguide Neural Operator (WGNO). This hybrid approach replaces the linear system solver with a Multilayer Perceptron (MLP) trained to approximate the solution. The WGNO is physics-informed, trained to minimize the residual of the physical laws (the linear system equation), ensuring the output satisfies the underlying physics.

C. Optimization Strategies

The framework optimizes the absorber permittivity $\varepsilon(x)$ using two reparameterization strategies, both followed by hard binarization:

Pixel-Wise Density: An unconstrained auxiliary field is mapped to physical density via a sigmoid function. The loss function includes terms for field matching, binarization penalty, and total variation (to suppress subwavelength artifacts).
Fourier-Parameterized Projection: The density is represented by a band-limited latent function defined by Fourier coefficients. This approach naturally suppresses high-frequency noise and enforces smooth mask walls. The loss function replaces the total variation penalty with a spectral penalty on high-order coefficients.

The optimization minimizes a loss functional measuring the discrepancy between the squared magnitudes of the reflected and desired electric fields on a remote surface.

3. Key Contributions

End-to-End Differentiable Framework: The paper presents a novel ILT framework where the full forward diffraction model (based on the rigorous waveguide method) is differentiable, enabling direct gradient-based optimization of mask permittivity via automatic differentiation.
Hybrid Physics-Informed Optimization: The integration of the WGNO into the ILT pipeline offers a hybrid scheme that retains the physics-based structure of the waveguide method while replacing the most expensive linear algebra step with a neural network.
Reparameterization Techniques: The study systematically compares pixel-wise density optimization against Fourier-parameterized projection, demonstrating the latter's ability to produce smoother, more manufacturable mask profiles with reduced optimization time.
Material and Dimensional Scalability: The framework is validated on realistic 2D and 3D absorbers (TaBN, Lanthanum, Uranium) at $\lambda=11.2$ nm, demonstrating scalability to multi-feature layouts and three-dimensional mask geometries.

4. Experimental Results

Numerical experiments were conducted on a 2D EUV mask setup with a 60 nm absorber layer and a 30-period Ru/Be/Sr mirror stack at $\lambda=11.2$ nm.

Performance of Materials: Three absorber materials were tested. Lanthanum (La) provided the best central maximum intensity, while Uranium (U) yielded the closest match to the desired field distribution. TaBN served as a baseline.
Optimization Efficiency: The Fourier-parameterized strategy achieved a 1.31-fold speedup (137 seconds vs. 179 seconds) compared to the pixel-wise method while solving the same forward problem. It also produced masks with well-defined walls and fewer jagged boundaries.
WGNO Performance: While the WGNO produced comparable optimization results to the rigorous solver, it did not currently offer a speed advantage. The authors note that the computational cost of training the network simultaneously with mask optimization outweighed the benefits of the faster forward solve in their current setup.
3D Extension: The framework successfully extended to 3D mask optimization (214×214 nm target pattern), confirming that the pipeline scales to realistic, complex layouts.

5. Significance and Claims

The paper claims to establish a fully differentiable route to EUV inverse mask design where the rigorous forward model is preserved throughout the optimization process. By leveraging automatic differentiation on the waveguide method, the authors demonstrate that it is possible to recover mask structures that achieve desired field distributions without relying on supervised learning datasets or gradient-free heuristics.

The authors position this work as a step toward overcoming the computational limitations of conventional ILT. They suggest that while the WGNO integration in its current form (simultaneous training) does not yet yield a net speedup, the framework is robust and scalable. The paper concludes that this approach is applicable not only to EUV mask design but also to the inverse design of metamaterials and other periodic electromagnetic structures, provided the forward model can be made differentiable.