Global-in-time Spectral Duhamel-Picard Method for Post-Exposure Bake Simulation and Physics-Informed Neural Operators for EUV Lithography

summary

The gist

The gist: The proposed method presents a global-in-time spectral Duhamel–Picard method and a physics-informed neural operator (PINO) that enable accelerated, time-marching-free prediction of

In short

The method introduces a global-in-time spectral Duhamel–Picard approach to simulate complex chemical reactions during Post-Exposure Bake (PEB) in lithography. It reformulates reaction-diffusion equations into integral equations solved via polynomial expansions and Anderson acceleration, avoiding slow sequential time marching. This is extended into a Physics-Informed Neural Operator (PINO) for fast, time-marching-free prediction of concentration fields.

Key concepts

Duhamel Formulation: This technique transforms the complex reaction-diffusion PDEs governing PEB into a form that incorporates linear diffusion through a heat semigroup. This allows the nonlinear acid and quencher trajectories to be represented as Volterra integral equations, which are then solved efficiently using global methods.
Global-in-Time Spectral Duhamel–Picard Method: This is the core numerical solver. It uses polynomial expansions to represent full concentration trajectories over the entire bake interval. Temporal coefficients are found using Anderson-accelerated Picard iteration, which solves coupled nonlinear Volterra integral equations without needing to march through time step-by-step.
Physics-Informed Neural Operator (PINO): The PINO is a machine learning model trained to predict the temporal coefficient fields rather than final concentration values at specific times. It uses physics constraints (PDE residuals and integral consistency losses) during training to ensure its predictions accurately represent the underlying physical process.
Mixed-transform f-FNO: This is a specific neural operator architecture used for the PINO. It replaces standard 3D spectral kernels with an axial operator followed by factorized lateral operators. This structure helps in efficiently learning how temporal coefficient fields evolve across the entire bake interval.

Terminology used across episodes

Reinforcement Learning (RL) — explained in 25 episodes
Sample Complexity — explained in 20 episodes
Retrieval-Augmented Generation (RAG) — explained in 18 episodes
Federated Learning — explained in 17 episodes
Quantum Fisher Information (QFI) — explained in 16 episodes
Flow Matching — explained in 15 episodes

This episode discusses

Global-in-time Spectral Duhamel-Picard Method for Post-Exposure Bake Simulation and Physics-Informed Neural Operators for EUV Lithography · Paper Radio

The paper

Global-in-time Spectral Duhamel-Picard Method for Post-Exposure Bake Simulation and Physics-Informed Neural Operators for EUV Lithography · Read on arXiv

Doyun Kim, Mircea Dusa, Werner Gillijns, Philippe Leray

imec

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: Today's paper: "Global-in-time Spectral Duhamel-Picard Method for Post-Exposure Bake Simulation and Physics-Informed Neural Operators for EUV Lithography".

Jane: The gist: The proposed method presents a global-in-time spectral Duhamel–Picard method and a physics-informed neural operator (PINO) that enable accelerated,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we've covered the core claims of this work on the "Global-in-time Spectral Duhamel–Picard Method for Post-Exposure Bake Simulation and Physics-Informed Neural Operators for EUV Lithography," which is essentially about using integral equations and neural operators to speed up PEB simulation.

Jane: They showed that this method allows for the accurate and efficient evaluation of the acid, quencher, protected group, and deprotected group concentration trajectories over the entire bake interval.

Lu: The core contribution is combining spectral Duhamel–Picard with a PINO trained on mixed-transform f-FNO to predict temporal coefficient fields.

Meng: The practical implication for production is that you can get faster predictions for these lithography simulations by learning the underlying temporal structure of the solution rather than calculating every single time step explicitly.

Tom: They demonstrated that this PINO preserves the principal three-dimensional post-development topology across randomly sampled mask layouts and PEB times while reducing the runtime relative to the global-in-time Duhamel–Picard method.

Jane: It means you get accurate predictions for where things are happening in the material, even near surfaces, which is important for process optimization.

Lu: The final point is that this PINO accurately and efficiently evaluates the acid, quencher, protected-group, and deprotected-group concentration trajectories over the entire PEB interval.

Tom: That's what it boils down to. This work provides a way to predict these complex concentration profiles much faster than traditional explicit time marching methods by using global integral formulations and neural operators.

Conclusion: Tom: So we've seen how they used spectral Duhamel–Picard methods to solve these reaction-diffusion equations, but now we're looking at the title of this paper itself: "Global-in-time Spectral Duhamel–Picard Method for Post-Exposure Bake Simulation and Physics-Informed Neural Operators for EUV Lithography."

Jane: That sounds super technical. What does that actually mean for someone who doesn't work on lithography?

Lu: It means they’ve got a method that tracks the acid and quencher reactions across the *entire* bake time, not just one instant, using these spectral math tools.

Meng: So it's trying to avoid those slow sequential time marching steps we usually have to do in simulations. Is that what they claim?

Tom: They claim they can predict concentration fields at any arbitrary time within the bake interval without having to march through every single millisecond explicitly, which is a big deal for speed.

Jane: So instead of running a long simulation just to see what happens at time T plus one, you can jump straight to it?

Lalam: That’s right. The method lets you look up the solution instantly because they used global polynomial expansions and then fed those into this neural operator structure.

Lu: The PINO part is key here; they aren't just solving the PDEs directly; they are training an AI to learn how those temporal coefficients evolve over the bake interval.

Tom: That means the AI learns the "shape" of time itself, which is way more efficient than calculating every single point in time one by one.

Jane: And what about accuracy? If you use this instead of a traditional solver, how good are these predictions going to be for things like acid diffusion?

Meng: The paper says they got a self-consistency level around ten to the negative seven after ten iterations, which suggests it’s pretty tight when it works.

Tom: It's about getting that accuracy using this combined integral formulation and the PINO extension, which is what makes the whole thing work together.

Jane: So, for someone who only listens to this show, what does this mean for how we think about complex chemical processes in microchips?

More episodes

← Home