euvsimulator

An open-source EUV lithography simulator whose every physical default has a named source.

euvsimulator models the EUV imaging and resist chain at 13.5 nm — multilayer mirror, mask diffraction, partially coherent imaging, chemically amplified resist, development, and the stochastics behind line-edge roughness — in Python and PyTorch, on a laptop.

What sets it apart is not the number of modules but the bookkeeping: each default of the resist chain is either measured in a cited paper, derived from a cited measurement, or explicitly labelled as a calibration or an assumption. Two named resists serve as anchors, and the distance between simulation and measurement is documented rather than tuned away. See docs/physics.md for the chain, the sources and the open questions.

What it is — and what it is not

• It is a transparent, tested physics chain for studying how exposure, bake, development and noise interact, with the provenance of every number one click away.
• It is not a predictive tool for production resists. At the NXE1716 anchor the printing dose comes out 1.7–1.9× above the measurement (an inconsistency inside the source data) and the roughness is bracketed between 0.7× and 2.4× by a blur nobody has measured for that resist; at the MET-2D anchor the simulated roughness is 4.3 nm 3σ against 6.7 nm measured with SEM bias. Two physics questions that no freely available measurement decides change results by a factor of two; euv calibrate --bands reports them as bands.

Quick start

pip install git+https://github.com/Flowbudget/euvsimulator.git

euv simulate --period 64 --cd 32 --dose 1.3 --resist-model full_chem   # default resist
euv process-window --period 64 --cd 32                                   # Bossung, DoF, EL
euv calibrate wafer.csv --period 44 --cd 22 --bands                      # fit your own FEM
euv serve                                                                # browser GUI + REST API

from euvsimulator.pipeline import SimulationConfig, run_simulation

cfg = SimulationConfig(period_nm=64, line_width_nm=32, dose_mj_cm2=1.3, resist_model="full_chem")
result = run_simulation(cfg)
print(f"CD {result.cd_nm:.1f} nm, NILS {result.nils_value:.2f}")

The default resist is a 2011 research resist (Yamamoto et al.) and prints at about 1.3 mJ/cm² at 64 nm pitch — far more sensitive than a production resist. Use the presets in euvsimulator.presets (NXE1716, MET-2D) or calibrate to your own data.

Browser GUI

euv serve starts a local server with a single-page GUI at http://localhost:8000/ and the REST API (/docs). No external assets, works offline.

• Presets with provenance — the documented default resist, the NXE1716 anchor as a prediction (dose-to-size 19.7 vs measured 11.0 mJ/cm²), the same anchor with the ×1.79 dose calibration labelled as such, and the MET-2D anchor. "Where these numbers come from" is one click away, and every result carries its notes.
• Every parameter — six main ones open, the other 58 SimulationConfig fields in collapsed groups with the same help texts as the CLI. Physical limits only (NA < 1, CD < pitch, …); grid, rows, seeds and realisations are unbounded, a live estimate says what a run will need in memory (model fitted on an M1, ±30 %).
• Three tasks — a single simulation (profile, CD, NILS, LER/LWR with n_eff and correlation length), a process window (Bossung table, depth of focus, exposure latitude) and the structural uncertainty bands of euv calibrate --bands. Runs are background jobs with progress and a cancel button; results export as CSV and JSON.
• Links reproduce runs — ?preset=met2d&task=process_window&f.grid=128&run=1 starts a run on load, ?job=<id> reopens a finished one.

What is inside

Stage	Model	Notes
Materials	CXRO atomic scattering factors, Z = 1–92	absorption of a resist computed from its composition
Multilayer mirror	S-matrix transfer-matrix method, Névot–Croce roughness	verified against Fresnel and energy-conservation limits
Mask	thin-mask Fourier series; RCWA 1D/2D optional	no flare, no mask roughness
Imaging	Abbe/Hopkins partially coherent, any NA < 1 with an isomorphic pupil	conventional, annular, dipole and quasar sources incl. scanner-style sector poles; an anamorphic high-NA pupil exists in aerial/pupil.py but is not wired into the pipeline
Exposure	Dill ABC with depth attenuation; photon shot noise with secondary-electron spread; PAG counting	acid yield per absorbed photon checked against measurements
Post-exposure bake	diffuse–quench–deprotect with measured kinetics at 80–140 °C; concurrent reaction-diffusion (NIST law) optional	validated on the NIST bilayer diffusion lengths
Development	Mack rate, Eikonal first-arrival front, sub-pixel CD; dissolution-cell noise optional	roughness metrology passband and SEM-bias bookkeeping
Metrology	CD, NILS, process window, LER/LWR with correlation-aware statistics
Calibration	fit to wafer FEM data, bootstrap CIs, structural uncertainty bands

Six executable notebooks in notebooks/ walk through the chain.

Validation status

Numerics are checked against exact solutions and conservation laws. The physics is anchored to primary measurements: Yamamoto 2011 (deprotection kinetics at seven temperatures, dissolution threshold), Kang/NIST 2009 (bilayer diffusion lengths), LBNL film quantum yields, Thackeray 2010 (blur decomposition), Vesters 2017/2019 (NXE1716 dissolution curves and 22 nm lines on an NXE3300B), Anderson & Naulleau 2008 (MET-2D blur, dose and LER). What the chain reproduces, and where it does not, is tabulated in docs/physics.md §6; the detailed record is docs/claude_code_arbeitslog.md.

Known limitations

• Not validated for production resists; see above.
• Undecided physics (factor 2): the acid-loss law in patterns (peb_model) and the size of the dissolving unit (dissolution_cell_nm). Experimental options are off by default and listed in the SimulationConfig docstring.
• Optics: 1D gratings, no flare, no mask roughness, defocus only. The plasma-source module is a parametric power-budget estimate and is not connected to the imaging chain.
• Compute: the stochastic chain at 2048 rows and grid 256 needs about 1.8 GB and minutes per realisation; the concurrent PEB costs ~6× more. Developed and tested on an Apple M1 with 8 GB.
• Tests run in CI on Linux, macOS and Windows with Python 3.10–3.13 (see the badge); numeric goldens were measured on an Apple M1 and hold on x86 to 1e-4 nm.
• The fast aerial_threshold model sets its threshold relative to the mean intensity of the image. That is fine for dose and focus sweeps of one mask, but when the mask itself changes the threshold moves with it: MEEF at 64/32 nm comes out 0.50 instead of 1.02 with a fixed threshold (1.28 in the full chemistry). Use full_chem for MEEF and mask comparisons.

Documentation

• docs/physics.md — the chain, every default with source and status, anchors, open questions
• CHANGELOG.md — release notes; docs/analyse_2026-09-05_stand_und_plan.md — plan and roadmap
• Sphinx sources in docs/ (pip install sphinx sphinx-rtd-theme && cd docs && make html)

Citation

@software{euvsimulator2026,
  author  = {Florian Scheide},
  title   = {euvsimulator: Open Source EUV Lithography Simulator},
  version = {2.2.1},
  year    = {2026},
  url     = {https://github.com/Flowbudget/euvsimulator},
  license = {Apache-2.0}
}

License and third-party material

Apache 2.0 — see LICENSE. All runtime dependencies are OSI-licensed (numpy, scipy, torch, matplotlib, fastapi, uvicorn, pydantic, typer, httpx, gdstk); the browser GUI has no external assets. CXRO scattering-factor tables are redistributed unmodified with attribution; anchor data are digitised numbers with their citations. See NOTICE.

Contributing and support

Issues and pull requests are welcome — see CONTRIBUTING.md. Physics changes follow the project rule: predict first, measure, then change the code, and keep the source of every number.

euvsimulator is developed and tested on a single laptop. If it saves you time, you can support the work: