Technology / CV Architecture
CNN architecture designed for semiconductor inspection
Not a fine-tuned general-purpose model. A convolutional neural network trained specifically on optical physics-based synthetic semiconductor wafer images.
Architecture
CNN inference pipeline
The classification model processes 64×64px patches extracted from the optical inspection scan. Inference runs at 28ms per patch, enabling 120 wph throughput on 300mm wafers.
Training Data
Solving the labeled data shortage
Real wafer defect images are scarce, expensive, and often proprietary. Lenspathio's training approach uses optical physics simulation to generate synthetic wafer defect images — calibrated to the specific illumination characteristics of each process node.
Optical physics-based augmentation
Defect images are synthesized by modeling defect geometry (particle shape, scratch depth, crystal pit morphology) under the specific illumination wavelength and NA configuration of the inspection system. Not generic image augmentation — physics-grounded signal generation.
Process node calibration
Each process node requires different training data because defect morphology and optical contrast change with feature pitch. The 3nm node training set includes synthetic data generated under 193nm ArF illumination — matching the actual inspection conditions.
Real wafer validation
Synthetic training data is validated against real wafer inspection images at each supported node. The false positive rate specification (<0.1%) is measured on real wafer data — not held-out synthetic data — before each model release.
No customer data required
The model ships trained. Customer wafer images are not used for training or model improvement — they stay on-premises. No participation in any federated learning scheme. Model updates are developed independently at Lenspathio.
Performance
Inference benchmarks
| Process Node | FPR (measured) | FNR (measured) | Inference Latency | Throughput |
|---|---|---|---|---|
| 28nm | <0.05% | 0.12% | 22ms | 120 wph |
| 16nm | <0.06% | 0.15% | 24ms | 120 wph |
| 10nm | <0.08% | 0.18% | 26ms | 120 wph |
| 7nm | <0.09% | 0.21% | 27ms | 120 wph |
| 5nm | <0.09% | 0.24% | 28ms | 118 wph |
| 3nm | <0.10% | 0.29% | 28ms | 115 wph |
FPR and FNR measured on held-out real wafer datasets at each process node. No external benchmark source cited — these are Lenspathio internal measurements.
See the architecture in action.
Request an evaluation and we run the full pipeline on your wafer data. You receive inference results, confidence score distributions, and false positive count at your specific process node.