PADE: Process-Adaptive Denormalization for Focus–Dose Conditioned SEM Image Generation

Code for the paper

H. Lee, S.-W. Lee, D.-N. Kim, "PADE: Process-Adaptive Denormalization for Focus–Dose Conditioned SEM Image Generation," submitted to Journal of Micro/Nanopatterning, Materials, and Metrology (JM3).

PADE generates SEM images of a layout at arbitrary focus and dose from a small set of anchor conditions. The generator is built on CLADE (class-adaptive normalization). In every normalization block, PADE maps the normalized process condition θ = (dose, focus) through a small MLP to a channel-wise scale and shift of the class-indexed affine parameters, so the process condition acts on every block of the generator. A VAE encoder provides a latent code z for the layout-specific appearance. At inference, z at a target condition is the bilinear interpolation of the latent means of the anchor images of the same cell.

Repository layout

train.py                     training (CLADE generator + PADE, optional VAE encoder)
generate.py                  generation at arbitrary focus-dose conditions (latent interpolation between anchors)
evaluation/cd_measure.py     row-wise CD and LWR measurement of line-space images (Appendix B)
evaluation/evaluate_cd_el.py CD agreement on held-out conditions and exposure latitude (EL) from Bossung surfaces
scripts/                     commands used for the experiments of the paper
models/ data/ options/ trainers/ util/   network, dataset, and option code

PADE itself is implemented in models/networks/normalization.py (ClassAffine, option --use_pade); the normalization of θ is in data/pix2pix_dataset.py (normalize_theta).

Installation

pip install -r requirements.txt

Tested with Python 3.12, PyTorch 2.11 and one NVIDIA GPU.

Data format

Masks and SEM images are 8-bit PNG files with matching names <cell>_f{focus}_d{dose}.png (focus in nm, may be negative; dose in the units of the simulator). The mask stores one class index per pixel.

• Training: --label_dir holds the mask of every training image (one copy per condition) and --image_dir the SEM images, here only the anchor conditions.
• Generation and evaluation: one mask per cell, <cell>.png, and the SEM images of the cell at the anchor conditions (and, for evaluation, at all conditions), either in one folder or in <cell>/ subfolders.

The process condition is parsed from the file names and normalized according to --theta_type:

--theta_type	study	θ
dose_focus	line-space focus-dose study (Sections 3.2–3.4)	((dose − 40)/20, focus/45)
dose_focus2d	additional layout families (Section 3.5)	((dose − 40)/20, (focus − 20)/50)
dose	single-variable dose study (Section 3.1)	((dose − 20)/20)

The simulated SEM dataset and the trained checkpoints are available from the corresponding author upon reasonable request.

Usage

Train the full model on the 3 × 3 anchor conditions (focus −45/0/45 nm × dose 40/50/60):

DATA=./datasets/focus_dose_anchors bash scripts/train_focus_dose.sh

Generate the dense 7 × 11 focus-dose grid and compare CD and EL with the simulation:

CKPT=./checkpoints/pade_focus_dose MASK=./datasets/mask GRID=./datasets/focus_dose_grid \
    bash scripts/generate_and_evaluate.sh

evaluate_cd_el.py prints the CD correlation, RMSE, and bias over the held-out (non-anchor) conditions and the mean absolute EL error and EL correlation over all cells. The ablation models, the single-variable study, and the additional layout families are trained with the other scripts in scripts/.

Acknowledgments and license

This code is built on CLADE (Tan et al.) and SPADE (Park et al.). Like these projects, it is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license (see LICENSE.md).