NanoFab Optimizer

NanoFab Optimizer is a small research prototype for exploring binary illumination patterns for nanoscale optical fabrication. It generates 7×7 digital micromirror device (DMD) motifs, simulates their diffraction-limited projection, and trains a machine-learning surrogate to predict their peak intensity.

The current model is intentionally simplified. It is useful for investigating pattern geometry and optimization workflows, but its results are not yet validated fabrication recipes.

What the project does

The repository implements this workflow:

1. Reduce a 7×7 binary pattern to 13 independent variables by requiring C4 symmetry (invariance under 90-degree rotations).
2. Enumerate all 2^13 = 8,192 symmetric motifs.
3. Project each motif through a coherent Fourier-optics model with an ideal circular pupil.
4. Measure peak intensity, horizontal full width at half maximum (FWHM), total energy, and active-pixel count.
5. Train a random-forest surrogate that predicts peak intensity from motif geometry.

The default optical parameters are:

Parameter	Value
Wavelength	804 nm
Numerical aperture	1.25
Projected pixel pitch	114 nm
Simulation grid	128×128

Repository layout

src/nanofab_optimizer/
  motifs.py                 C4 motif generation and 13-bit encoding
  optics.py                 Coherent diffraction simulation and FWHM measurement
scripts/
  run_motif.py              Simulate and plot one example motif
  generate_motif_dataset.py Simulate all 8,192 motifs
  analyze_motif_dataset.py  Summarize and visualize the dataset
  train_peak_surrogate.py   Train and evaluate random-forest models
tests/                      Unit tests for motifs and optics
outputs/                    Generated dataset and plots

Installation

The project requires Python 3.11 or newer and uses uv for environment and dependency management.

git clone <repository-url>
cd nanofab-optimizer
uv sync

To include the test and lint dependencies:

uv sync --group dev

Usage

Run a single example motif and create outputs/motif_v01.png:

uv run python scripts/run_motif.py

Generate the complete C4-symmetric dataset:

uv run python scripts/generate_motif_dataset.py

This writes outputs/motif_dataset_v02.npz. Analyze it and regenerate the summary figures with:

uv run python scripts/analyze_motif_dataset.py

Train and evaluate the peak-intensity surrogate:

uv run python scripts/train_peak_surrogate.py

Run the test suite:

uv run pytest

Current results

The checked-in dataset contains all 8,192 C4-symmetric motifs. Relative values use the fully illuminated 7×7 motif as the reference.

Result	Value
Highest relative peak intensity	1.1144
FWHM of highest-peak motif	440.5 nm
Active pixels in highest-peak motif	37 of 49
All-on reference FWHM	531.6 nm
Narrowest measured FWHM	330.4 nm
Relative peak of narrowest motif	0.0052

The highest-peak motif is:

0 0 1 1 1 0 0
0 1 1 1 1 1 0
1 1 1 1 1 1 1
1 1 1 1 1 1 1
1 1 1 1 1 1 1
0 1 1 1 1 1 0
0 0 1 1 1 0 0

On the fixed 80/20 train/test split, the geometry-aware random forest achieves an R² of approximately 0.998 and identifies the true highest-peak motif. A baseline using only the number of active pixels achieves an R² of approximately 0.42, showing that spatial arrangement matters substantially.

Optical model

simulate_motif treats the binary motif as a coherent input field. It centers that field on the simulation grid, transforms it into spatial-frequency space, removes frequencies above NA / wavelength, and inverse-transforms the filtered field. Intensity is the squared magnitude of the resulting complex field.

FWHM is measured along the horizontal profile through the global intensity maximum, with linear interpolation at the half-maximum crossings.

Limitations

The simulator is a V0.1 coherent diffraction model, not a complete GP-TPL or fabrication model. It currently omits effects such as:

• broadband femtosecond illumination;
• nonlinear two-photon material response and exposure thresholds;
• aberrations, polarization, and DMD phase effects;
• defocus and full 3D propagation;
• process constraints and experimental calibration.

The surrogate is also evaluated on samples drawn from the same finite design space used for training. Since the current 8,192-pattern space can be enumerated directly, its main purpose is to demonstrate an approach that can scale to larger, non-exhaustive searches.

Development

Format and lint the project with Ruff:

uv run ruff check .
uv run ruff format --check .

The reusable code lives under src/nanofab_optimizer; the scripts currently act as direct, configuration-by-source research experiments.