🇬🇧 An integrated AI solution for semiconductor wafer edge defect detection.
🇨🇳 一体化半导体晶圆边缘缺陷检测 AI 解决方案。
EN: This project targets the hardest problem in semiconductor manufacturing — detecting wafer edge defects (chipping, particles, scratches, dislocations) on ARM edge devices with only 30–50 labeled samples. The core innovation is a shared-encoder dual-branch architecture (RepViT backbone) with optical physics prior injection (PISM + ASG) — achieving < 80ms inference per frame on ARM Cortex-A76, INT8 model < 8MB.
ZH: 本项目解决半导体制造中最前沿的难题——在 ARM 边缘设备上仅用 30-50 张标注样本实现晶圆边缘缺陷检测（崩边、颗粒、划痕、位错）。核心创新为共享编码器双分支架构（RepViT 骨干）结合光学物理先验注入（PISM + ASG）—— ARM Cortex-A76 上单帧推理 < 80ms，INT8 模型 < 8MB。
暗场灰度图 (512×512)
│
▼
┌───────────────────────────┐
│ RepViT-M0.9 共享编码器 │
└────────┬─────────┬────────┘
│ │
┌────────▼┐ ┌─────▼──────────┐
│ 增强解码 │ │ 检测输出分支 │
│ (U型降噪) │ │ Anchor-free │
│ → 明场图 │ │ PISM+ASG+SGF │
└─────────┘ └──────┬──────────┘
│
┌───────────────────▼────────────┐
│ TorchScript JIT + INT8 量化 │
│ → ARM C++ SDK < 80ms/帧 │
└────────────────────────────────┘
| Service | Stack | Role |
|---|---|---|
| wafer-inspection | PyTorch | Core detection model + training |
| wafer-backend | Spring Boot 3.2 | API gateway, business logic |
| wafer-lora-service | FastAPI + SD 1.5 | LoRA-based image generation |
| wafer-qwen-service | FastAPI + llama.cpp | Qwen-VL defect analysis |
| wafer-showcase | React 18 + Vite 5 | Frontend dashboard |
git clone https://github.com/your-org/wafer-inspection.git
cd wafer-inspection
docker-compose up -d
# Visit http://localhost:80See individual service READMEs for detailed instructions.
| Module | Params | Function |
|---|---|---|
| PISM | 0.16M | Differentiable Rayleigh-Mie scattering: virtual 266nm→193nm |
| ASG | 0.01M | 13-angle scatter generation via torch.roll (zero-cost) |
| Dual-branch | 0.37M | 266nm (precision) + virtual 193nm (recall) parallel detection |
| SGF | 0.008M | Scattering-guided fusion for wavelength-aware merging |
Insight: Replace a $500K 193nm DUV light source with AI — Rayleigh gain (266/193)⁴ ≈ 3.5× for small defects.
50 labeled samples → Copy-Paste + Mosaic + MixUp → 10,000+ effective samples
↓
Stage 1: CycleGAN unsupervised pretraining (0 labels needed)
Stage 2: Joint fine-tuning (50 labels)
Stage 3: Knowledge distillation (7.5M → 4.2M, < 1% accuracy loss)
| Constraint | Target | Solution |
|---|---|---|
| Parameters | < 10M | RepViT-M0.9 + reparameterization |
| INT8 model | < 8MB | Per-channel quantization + sensitive-layer FP32 skip |
| ARM latency | < 80ms/frame | Polar transform + reparameterization fusion |
| 12" wafer | < 3min | 2000-4000 sub-images batched pipeline |
| Doc | 说明 |
|---|---|
| Technical Reference | 完整技术路线（压缩版） |
| Optical Physics | 光学物理约束注入详解 |
| Project Manager Guide | 非技术负责人导读 |
| Deployment Guide | 部署运维指南 |
| Qwen-VL Detection | Qwen-VL 晶圆检测方案 |
| Showcase Design | 前端展厅设计文档 |
| Competition Guide | 光电竞赛策划指导 |
MIT License — see LICENSE for details.
- School: 东北大学秦皇岛分校 (Northeastern University at Qinhuangdao)
- Project: 晶圆缺陷检测·一体化 AI 解决方案
- Year: 2026
全链路自主可控 · 不依赖进口光学硬件 · 算法全栈自主
