man-svi/Drift-Sense-FinFET-SEM

Drift-Sense Synthetic FinFET SEM Dataset

Overview

The Drift-Sense Synthetic FinFET SEM Dataset is a synthetic computer-vision dataset developed for the Drift-Sense project.

The project investigates navigation and localization errors that can occur when a semiconductor inspection system attempts to relocate a previously identified site within a larger search region.

The dataset provides paired high-resolution reference images and larger search images containing the corresponding target region.

The dataset is designed for controlled experimentation in:

• Image localization
• Template matching
• Robust computer vision
• Navigation-error detection
• Localization recovery
• Semiconductor inspection research

Important: This is a synthetic FinFET-style dataset. It is not collected from a commercial SEM or a specific semiconductor manufacturing process.

---

Dataset Summary

Property	Value
Architecture style	FinFET
Total image pairs	2,400
Training pairs	2,000
Validation pairs	200
Test pairs	200
Total PNG images	4,800
Reference resolution	1000 × 1000 pixels
Search resolution	1000 × 1000 pixels
Reference sampling	1 nm/pixel
Search sampling	10 nm/pixel
Spatial scale ratio	10×
Target patch at search scale	100 × 100 pixels
Image type	Grayscale

---

Dataset Structure

The repository is organized as follows:

dataset/
├── train/
│   ├── reference/
│   └── search/
│
├── val/
│   ├── reference/
│   └── search/
│
└── test/
    ├── reference/
    └── search/

metadata/
├── train.csv
├── val.csv
└── test.csv

Each sample contains:

1.A reference image representing the target site at high resolution.
2.A search image representing a larger inspection region.
3.Metadata containing the ground-truth target location.

Image Specifications
1.Reference Image
The reference image represents a high-resolution local region.
Resolution: 1000 × 1000 pixels
Sampling: 1 nm/pixel
Physical field of view: 1 µm × 1 µm

2.Search Image
The search image represents a larger inspection region.
Resolution: 1000 × 1000 pixels
Sampling: 10 nm/pixel
Physical field of view: 10 µm × 10 µm
10× Scale Relationship

The reference and search images use a 10× spatial sampling
relationship.

The 1000 × 1000 pixel reference image corresponds to a
100 × 100 pixel target region when represented at the search-image
resolution.

Reference
1000 × 1000 px
1 nm/pixel
      │
      │ 10× spatial scale
      ▼
Target at search resolution
100 × 100 px
10 nm/pixel

Synthetic FinFET-Style Structures

The reference images are procedurally generated using simplified
FinFET-style structural patterns.

The generator includes:
1.Parallel fin structures
2.Periodic gate structures
3.Contact features
4.Repeated local patterns
5.Intensity variations

The search background contains corresponding large-scale periodic
structures so that the localization problem contains repeated and
potentially ambiguous patterns.

These structures are intended to provide a controlled environment
for studying localization errors rather than to reproduce a specific
fabrication process.

Image Degradation and Variations

To evaluate robustness, controlled imaging variations are introduced
during dataset generation.

These include:

Gaussian noise
Poisson noise
Blur
Contrast and brightness variation
Edge enhancement
Charging-like intensity artifacts

The reference and search images are independently augmented.

These variations are intended to represent challenging imaging
conditions that can affect feature matching and localization.

Metadata

Metadata is provided separately for each dataset split:

metadata/train.csv
metadata/val.csv
metadata/test.csv

Important metadata fields include:

| Field | Description |
|---|---|
| `sample_id` | Unique sample identifier |
| `architecture` | Architecture type |
| `reference_path` | Reference image path |
| `search_path` | Search image path |
| `reference_size` | Reference image dimensions |
| `search_size` | Search image dimensions |
| `reference_nm_per_pixel` | Reference spatial sampling |
| `search_nm_per_pixel` | Search spatial sampling |
| `scale` | Spatial scale ratio |
| `ground_truth_x` | Target center X-coordinate |
| `ground_truth_y` | Target center Y-coordinate |
| `patch_x` | Target top-left X-coordinate |
| `patch_y` | Target top-left Y-coordinate |
| `patch_width` | Target width |
| `patch_height` | Target height |

The ground-truth coordinates allow localization algorithms to be
evaluated quantitatively.

Dataset Splits

The dataset is divided into three independent splits:

Training:
    2,000 image pairs


Validation:
    200 image pairs


Testing:
    200 image pairs

The test split is intended for final evaluation of localization
performance.

Intended Use

This dataset is intended for research and educational experimentation
in:

Semiconductor image analysis
Wafer inspection algorithms
Image localization
Template matching
Computer vision
Robust localization
Navigation-error detection
Navigation-error recovery
Algorithm benchmarking

The dataset can be used to investigate how localization performance
changes under clean, noisy, blurred, geometrically challenging, and
combined imaging conditions.

Drift-Sense Project

The dataset was created as part of the Drift-Sense project.

Drift-Sense investigates a navigation-error recovery pipeline for
semiconductor inspection scenarios.

The overall experimental workflow includes:

Reference Image
       │
       ▼
Initial Localization
       │
       ▼
Confidence / Error Analysis
       │
       ▼
Difficult-Case Detection
       │
       ▼
Recovery Strategy
       │
       ▼
Recovered Localization

The dataset provides controlled ground-truth coordinates so that the
localization and recovery stages can be evaluated quantitatively.

Evaluation Metrics

Localization performance can be evaluated using positional error
between the predicted target location and the ground-truth location.

The project uses metrics including:

Accuracy within 1 pixel
Accuracy within 3 pixels
Accuracy within 5 pixels
Mean localization error
Median localization error
Maximum localization error
Inference time

These metrics allow performance to be compared across different
imaging conditions and localization methods.

Limitations

This dataset is synthetic.

The FinFET structures are simplified procedural representations and
should not be interpreted as physically exact simulations of:

A specific commercial SEM
A specific semiconductor fabrication process
A particular process node
A particular wafer inspection tool
Actual manufacturing data

The dataset is therefore intended primarily as a reproducible
algorithm-development and benchmarking resource.

Real semiconductor inspection systems may contain additional
physical effects, imaging characteristics, process variations, and
instrument-specific errors that are not represented here.