LithoNet

LithoNet is a learning-based lithography simulation and mask optimization toolkit. Considering the analogy between image translation and lithography computation flow, LithoNet attempts to replace the conventional Lithography Simulation and ILT (Inverse Lithography Technology) processes with GAN-based image inference and generation. With golden datasets produced by commercial software (e.g. Synopsys S-Litho) based on rigorous numerical calculation, LithoNet has the following features:

Both lithography simulation flow and ILT flow are implemented.
Much faster process and acceptable quality, compared with commercial tools.
EPE (Edge Placement Error)-aware training and EPE distribution measurement for all contacts.
Easy for customization with your own datasets and upgradation.
...

Example results

Target Layout	Synthesized Mask	Aerial Image	Actual Layout

Reference Flow (Lin, ASPDAC'21)

Prerequisites

Python3
Nvidia GPU and CUDA
Pytorch (1.4+) and other dependencies (e.g., torchvision, opencv)
- For pip users, please type the command pip install -r requirements.txt
- For Conda users, create a new Conda environment using conda env create -f environment.yml

Getting Started

Installation

Clone this repo:

git clone https://github.com/xinming-wei/LithoNet
cd LithoNet

Lithography Simulation

Input: mask.
Output: aerial image (intensity map).

Customize with Your Own Dataset

Put corresponding items in ./datasets/litho_simul/train/mask and ./datasets/litho_simul/train/aerial_image respectively for training, and similarly for testing in ./datasets/litho_simul/test/

Only one .npz file is needed in each subfolder, you can name it casually but its index should be like np.load('masks.npz')['0066']. Do it by:

import numpy as np
mask_dict = {}
for i, image in enumerate(images):
    mask_dict['{:04}'.format(i)] = image
np.savez('./datasets/litho_simul/train/mask/masks.npz', **mask_dict)

It's your job to keep the shape of images and the number of samples in paired datasets (e.g. masks and aerial images) aligned.
I have placed some example datasets with few images in the folders. You need to replace them with yours before training.

Train and Test

Train with the default 256x256 resolution:

#!./scripts/train_lithosimul.sh
python train.py --phase litho_simul

Test the model:

#!./scripts/test_lithosimul.sh
python test_lithosimul.py --plot

The test results will be saved as log files and images. You can find them in ./checkpoints/litho_simul/test_results/.

ILT Flow

Input: target layout.
Output: optimized mask (e.g. with OPC and SRAF).
Pretrained lithography simulation model is required to train ILT.

Pretrain

Like training lithography simulation model, you need to add your own dataset to ./datasets/ilt_pretrain/. Here aligned layout and mask datasets are required to pretrain the ILT generator:
```
#!./scripts/train_ilt_pretrain.sh
python train.py --phase pretrain
```

Finetune

At the finetune stage, the pretrained model is further refined with decreased learning rate and fewer epochs. In an unsupervised manner, only input target layout data is needed. Losses are computed based on the L2 error between target layout and actual wafer image.
To prepare the datasets, use the ./util/RandomClip/RandomClip.py to generate any number of layout images with your own configuration. Then put the .npz files under ./datasets/ilt_finetune/train and ./datasets/ilt_finetune/test.
You need to specify the threshold for slicing aerial image with continuous pixel value into binary printed layout (i.e. a canonical resist model). Use --aerial_thresh to set the value when training.

Train the finetune model:

#!./scripts/train_ilt_finetune.sh
python train.py --phase finetune --aerial_thresh [(float)your chosen value]

Test the final ILT model:

#!./scripts/test_ilt.sh
python test_ilt.py --plot --aerial_thresh [(float)your chosen value]

You can adjust the learning rate or training epochs by specifying --lr and --n_epochs, --n_epochs_decay. The test results (EPE per contact, test images and EPE distribution) will be saved in ./checkpoints/ilt_finetune/test_results/.

More Training/Testing Details

There are more options you can use (e.g. train on multiple GPUs, change batch size and epochs, switch to a higher resolution).
Check the details in ./options. see ./options/base_options.py and ./options/train_options.py for all the train flags; see ./options/base_options.py and ./options/test_options.py for all the test flags.

Acknowledgements

We would like to thank Synopsys, Inc. for supplying us with Sentaurus Lithography (S-Litho) to generate datasets and verify mask quality. This model also borrows heavily from pytorch-CycleGAN-and-pix2pix.
The tool still suffers from mask quality problems, so pull requests to improve the model are always welcome.