AISDL-SNU/LithoBench-PDE

---
license: cc-by-nc-4.0
task_categories:
- image-to-image
tags:
- computational lithography
- scientific machine learning
- PDE
pretty_name: LithoBench-PDE
size_categories:
- 1K<n<10K
---

# LithoBench-PDE

A benchmark dataset for PDE-based computational lithography simulation, constructed by generating high-fidelity 3D reference simulations for photomasks from the [LithoBench](https://github.com/shelljane/lithobench) dataset. Each sample contains intermediate 2D and 3D field data from the lithography simulation pipeline, providing ground-truth input-output pairs for three PDE learning tasks corresponding to three governing PDEs of photolithography.

## PDE Learning Tasks

For each photomask, the reference simulation pipeline generates three ground-truth input-output pairs:

| Task | PDE | Mapping | Input | Output | Shape |
|------|-----|---------|-------|--------|-------|
| Mask illumination | Maxwell's equation | M → E | `M` (photomask) | `E` (diffracted near field) | `[H, W]` → `[2, 2, H, W]` (complex64) |
| Post-exposure bake | Reaction-diffusion equation | h → m | `h` (photoacid concentration) | `m` (deprotection image) | `[25, H, W]` → `[25, H, W]` |
| Development | Eikonal equation | R → T | `R` (development rate) | `T` (development time) | `[25, H, W]` → `[25, H, W]` |

- **M → E**: Given a 2D photomask pattern, solve Maxwell's equations to predict the diffracted near field (DNF), represented as a 2×2 Jones matrix of the electric field (complex-valued). Reference solutions are computed by rigorous coupled wave analysis (RCWA).
- **h → m**: Given a 3D photoacid concentration volume (25 z-slices), solve the reaction-diffusion equation governing the deprotection reaction during post-exposure bake (PEB) to produce the deprotection image. Reference solutions are computed by the finite difference method (FDM).
- **R → T**: Given a 3D development rate field (25 z-slices), solve the eikonal equation to obtain the development time field, whose isosurface defines the 3D developed photoresist structure. Reference solutions are computed by the fast marching method (FMM).

All data are generated under a fixed nominal condition of an annular source, 0 nm focus, and 20 mJ/cm² dose, with a uniform grid spacing of 4 nm.

## Photomask Categories

| Category | Samples | Spatial Size | Description |
|----------|---------|-------------|-------------|
| Metal_I | 1,600 | 512×512 | Curvilinear metal photomasks (train) |
| Metal_T | 1,600 | 512×512 | Rectilinear metal target layouts (train) |
| Contact_I | 163 | 512×512 | Curvilinear contact photomasks (test, out-of-distribution) |
| Contact_T | 163 | 512×512 | Rectilinear contact target layouts (test, out-of-distribution) |

**Total: 3,526 samples (~378 GB)**

The photomask plane is represented on a 2048 nm × 2048 nm domain (512 × 512 grid at 4 nm spacing). 3D photoresist fields are represented on a 2048 nm × 2048 nm × 100 nm domain (512 × 512 × 25 grid). The DNF E is represented with four complex electric field components E_UV (U, V ∈ {x, y}), where E_UV denotes the U component of the electric field for incident light polarized in V direction.

## Download

### Option 1: Download zip archives (recommended for full categories)

Pre-packaged zip files are available for each category:

```python
from huggingface_hub import hf_hub_download

# Download a single category as zip
path = hf_hub_download(
"AISDL-SNU/LithoBench-PDE",
"zip/Contact_I.zip",
repo_type="dataset",
local_dir="./data",
)
```

```bash
# Or using the CLI
huggingface-cli download AISDL-SNU/LithoBench-PDE zip/Contact_I.zip --repo-type dataset --local-dir ./data
```

### Option 2: Download individual .pt files

```python
from huggingface_hub import hf_hub_download

path = hf_hub_download(
"AISDL-SNU/LithoBench-PDE",
"LithoBench_PDE/Contact_I/INV_X8__0_0.pt",
repo_type="dataset",
)
```

### Option 3: Download an entire category folder

```python
from huggingface_hub import snapshot_download

snapshot_download(
"AISDL-SNU/LithoBench-PDE",
repo_type="dataset",
allow_patterns="LithoBench_PDE/Contact_I/*",
local_dir="./data",
)
```

## Usage

### PyTorch Dataset

Download `LithoBench_PDE.py` from this repository and use the provided `LithoBenchPDE` Dataset class:

```python
from LithoBench_PDE import LithoBenchPDE

# Load from HuggingFace Hub — downloads .pt files to local cache
ds = LithoBenchPDE.from_hub("AISDL-SNU/LithoBench-PDE", categories=["Contact_I"])

# Load a specific PDE task (returns {"input", "target", "sample_id", "category"})
ds = LithoBenchPDE.from_hub("AISDL-SNU/LithoBench-PDE", task="maxwell")
ds = LithoBenchPDE.from_hub("AISDL-SNU/LithoBench-PDE", task="reaction_diffusion")
ds = LithoBenchPDE.from_hub("AISDL-SNU/LithoBench-PDE", task="eikonal")

# Use with DataLoader
from torch.utils.data import DataLoader
loader = DataLoader(ds, batch_size=4, shuffle=True)
for batch in loader:
inputs, targets = batch["input"], batch["target"]
```

```python
# Load from local directory (after downloading/unzipping)
ds = LithoBenchPDE("./LithoBench_PDE", categories=["Contact_I", "Metal_I"])
sample = ds[0]
# sample keys: M, E, h, m, R, T, sample_id, category
```

### Direct loading with PyTorch

```python
import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download(
"AISDL-SNU/LithoBench-PDE",
"LithoBench_PDE/Contact_I/INV_X8__0_0.pt",
repo_type="dataset",
)
sample = torch.load(path, map_location="cpu")

# Access fields
M = sample["M"] # [512, 512] float32 - photomask pattern
E = sample["E"] # [2, 2, 512, 512] complex64 - diffracted near field (Jones matrix)
h = sample["h"] # [25, 512, 512] float32 - photoacid concentration
m = sample["m"] # [25, 512, 512] float32 - deprotection image
R = sample["R"] # [25, 512, 512] float32 - development rate
T = sample["T"] # [25, 512, 512] float32 - development time
```

## File Format

Each `.pt` file is a Python dictionary saved with `torch.save()` containing:

```python
{
"M": torch.Tensor, # [H, W] float32 - photomask pattern
"E": torch.Tensor, # [2, 2, H, W] complex64 - diffracted near field
"h": torch.Tensor, # [25, H, W] float32 - photoacid concentration
"m": torch.Tensor, # [25, H, W] float32 - deprotection image
"R": torch.Tensor, # [25, H, W] float32 - development rate
"T": torch.Tensor, # [25, H, W] float32 - development time
}
```

## Citation

TBA