BiSMO: Bilevel Source Mask Optimization

Overview

BiSMO is a PyTorch-based framework for bilevel optimization in lithography applications, specifically for Source Mask Optimization (SMO). It provides a modular and efficient approach to solve the complex optimization problems in lithography by leveraging bilevel optimization techniques.

Installation

# Clone the repository
git clone https://github.com/yourusername/BiSMO.git
cd BiSMO

# Create and activate a conda environment (recommended)
conda create -n smo python=3.8
conda activate smo

# Install dependencies
pip install -r requirements/requirements.txt

Quick Start

BiSMO supports multiple bilevel optimization approaches:

• DARTS: Differentiable Architecture Search method
• NMN: Neumann Series method
• CG: Conjugate Gradient method
• RL: Reinforcement Learning approach

To run an experiment with one of these approaches:

# Using DARTS approach
./darts_d2.sh

# Using Neumann Series approach
./nmn_d1.sh

# Using Conjugate Gradient approach
./cg_d0.sh

# Using Reinforcement Learning approach
./rl_d1.sh

Structure

The project is organized as follows:

• src/: Main source code

  • bilevel.py: Entry point for bilevel optimization
  • betty/: Betty library integration
  • models/: Neural network models
    • mo_module.py: Mask Optimization module
    • so_module.py: Source Optimization module
  • problems/: Optimization problems
    • mo.py: Mask Optimization problem
    • so.py: Source Optimization problem
  • engine/: Optimization engines
  • data/: Data handling utilities

• configs/: Configuration files

  • problems/: Configuration for different optimization approaches
  • module/: Model configurations
  • engine/: Engine configurations

• Shell scripts (e.g., darts_d2.sh, nmn_d1.sh, etc.): For running experiments with different configurations

Configuration

BiSMO uses Hydra for configuration management. Main configuration options include:

• problem_type: Optimization approach (darts, nmn, cg, rl)
• unroll_steps: Number of unrolling steps in optimization
• device_id: GPU device to use
• Weights for different loss components
• Learning rates and other hyperparameters

Running Experiments

The provided shell scripts run optimization on different image masks with various approaches:

# Run optimization on 10 masks using DARTS approach on GPU 2
./darts_d2.sh

# Run with different unrolling steps (e.g., 3 steps) using Neumann Series
./nmn_d1_unroll3_iter.sh

# Run with different Conjugate Gradient iterations
./cg_d0_unroll3_iter.sh

Customization

You can modify the configuration files in configs/ to customize:

• Optimization approaches and hyperparameters
• Model architectures
• Input data sources and masks
• Training and validation settings