EUVLithography

A copy of the EUVlitho research code by Hiroyoshi Tanabe and colleagues (takahashi-edalab/EUVlitho, MIT), with a Streamlit viewer and a matplotlib analysis script added on top for exploring the simulator's precomputed output.

The electromagnetic simulator and the CNN are the reference implementation for a series of published papers (listed under Attribution). Layered on top are streamlit_app.py, analyze_data.py, setup_windows.py and the Windows launcher — a visualisation layer over the CSV files the upstream pipeline produces. The original upstream documentation is preserved verbatim at include/README.md and is the authoritative guide to the simulator itself.

What the upstream code does

EUV lithography masks are not flat. At 13.5 nm with a 60 nm-thick absorber, the mask behaves as a 3D scatterer, and the thin-mask (Kirchhoff) approximation — just Fourier transforming the layout — is visibly wrong. Modelling it properly means solving Maxwell's equations over the mask, which is slow.

The upstream pipeline has two halves:

1. Rigorous EM simulator (emint/intensity.cpp, include/header.h) — a 3D waveguide model, equivalent to RCWA but with fewer field components. The committed configuration is λ = 13.5 nm, NA = 0.33, 4× magnification, 6° chief ray angle, dipole illumination (σ 0.55–0.9, 90° opening), 60 nm absorber (n = 0.9567 + 0.0343i) over a 40-pair multilayer, on a 2048 nm pitch sampled at 512×512. It builds on Eigen + oneAPI MKL for the dense linear algebra and MAGMA/CUDA for the GPU eigensolves.
2. CNN surrogate (cnn/, cnnpredict/) — six PyTorch Lightning CNNs that predict the real and imaginary parts of the M3D parameters a0, ax, ay directly from the mask bitmap, skipping the EM solve. Each is five conv/BatchNorm/max-pool blocks (16→32→64→128→256 channels, 3×3, circular padding so the periodic mask wraps correctly) into a 4096-unit FC layer, regressing 1901 diffraction orders for a0 and 1749 for ax/ay under MSE loss. Training uses pattern-shift augmentation: a random cyclic roll of the mask with the matching phase ramp applied analytically to the targets, which expands 20,000 rigorous simulations into 1,000,000 training examples. Predicted M3D parameters are then fed to either Abbe summation (cnnabbe/) or the STCC/SOCS formulation (cnnsocs/) to form the aerial image.

Accuracy of the committed example

The repository ships one worked example — the same 512×512 mask run through every path. Comparing those CSVs directly (parse each file past its 4 header lines, take the array, subtract elementwise):

Aerial image	vs. rigorous EM	RMSE	Max abs. error
cnnabbe/nnabbe.csv — CNN + Abbe	emint/emint.csv	0.0061	0.0242
cnnsocs/nnsocs.csv — CNN + SOCS	emint/emint.csv	0.0059	0.0236
cnnabbe/ftint.csv — thin-mask FT	emint/emint.csv	0.0160	0.0392

Peak intensity in the EM reference is 0.657, so the CNN's RMSE is about 0.9% of peak, and roughly 2.6× closer to the rigorous result than the thin-mask approximation. That is the point of the method, and it is the one quantitative claim this repository can actually support — it comes from files that are committed here. It is a single pattern, not a benchmark over a test set.

What is missing

• No trained weights. No model.ckpt anywhere; the CNNs cannot be run.
• No training data. cnn/data/ampdata/*.npy and maskdata/mask.npy each hold a single sample; cnn/model/*/.py expect ndata=20000 / nval=5000 from absolute paths under /home/tanabe/..., which are not in the repo.
• The C++ makefiles (emint/makeint and friends) also carry the original author's hardcoded /home/tanabe/magma and Eigen paths and need editing before they build.

So the EM simulator and the CNN cannot be reproduced from this checkout alone. What runs is the visualisation of the already-computed CSVs.

Running the viewer

pip install -r streamlit_requirements.txt
streamlit run streamlit_app.py      # or: run_streamlit.bat on Windows
python analyze_data.py              # writes output_mask.png, output_intensity.png, output_comparison.png

Both read emint/mask.csv and emint/emint.csv and plot them. Note that the "Run CNN Inference" button in the demo tab is a placeholder — it sleeps briefly and prints a fixed message; no model is loaded.

Attribution

All C++ and CNN source here is from takahashi-edalab/EUVlitho, MIT License, Copyright (c) 2024 Hiroyoshi Tanabe. The method is described in:

• H. Tanabe, M. Shimode and A. Takahashi, "Rigorous electromagnetic simulator for extreme ultraviolet lithography and convolutional neural network reproducing electromagnetic simulations," JM3 24 (2025) 024201. https://doi.org/10.1117/1.JMM.24.2.024201
• H. Tanabe, A. Jinguji and A. Takahashi, "Weakly guiding approximation of a three dimensional waveguide model for extreme ultraviolet lithography simulation," JOSA A 41 (2024) 1491. https://doi.org/10.1364/JOSAA.516610
• H. Tanabe, S. Sato and A. Takahashi, "Fast EUV lithography simulation using convolutional neural network," JM3 20 (2021) 041202. https://doi.org/10.1117/1.JMM.20.4.041202
• H. Tanabe and A. Takahashi, "Data augmentation in extreme ultraviolet lithography simulation using convolutional neural network," JM3 21 (2022) 041602. https://doi.org/10.1117/1.JMM.21.4.041602
• H. Tanabe, A. Jinguji and A. Takahashi, "Evaluation of convolutional neural network for fast extreme ultraviolet lithography simulation using 3nm node mask patterns," JM3 22 (2023) 024201. https://doi.org/10.1117/1.JMM.22.2.024201
• H. Tanabe, A. Jinguji and A. Takahashi, "Accelerating extreme ultraviolet lithography simulation with weakly guiding approximation and source position dependent transmission cross coefficient formula," JM3 23 (2024) 014201. https://doi.org/10.1117/1.JMM.23.1.014201

The illumination model in ampS (include/header.h) follows N. Davydova et al., SPIE 88860A.