Technology

ML architecture purpose-built for cleanroom conditions

Convolutional inference designed for fab-floor electromagnetic interference, ambient vibration, and resist chemistry variation. Not adapted from general-purpose computer vision — built specifically for semiconductor defect classification.

Request Evaluation
Abstract representation of computer vision inspection technology — precision optical measurement in semiconductor manufacturing context
Technical Approach

Four design principles

Semiconductor-specific training data

Models trained on semiconductor defect image libraries — not general object detection adapted from ImageNet. Training includes controlled samples of each defect class at multiple process nodes and from multiple tool types.

Per-recipe adaptive calibration

Each recipe (process node + layer + tool) gets its own threshold profile derived from the first 50 production wafers. The calibration layer adjusts automatically as process conditions evolve. Keeps false-positive rates stable across tool lifetimes.

On-premises inference, no cloud

All ML inference runs on the local industrial compute unit. No wafer image data leaves your facility. No network dependency for inspection uptime. IP protection built into the deployment model.

EMI-hardened capture pipeline

Frame capture and pre-processing pipeline designed for fab-floor electromagnetic interference conditions. Optical sensor timing and capture sequencing validated against SEMI S22 equipment installation standards.

Architecture

Model + calibration layer separation

The base CNN model handles feature extraction — it learns what semiconductor defects look like in general. The per-recipe calibration layer handles the specific characteristics of your tool and process: optical aberrations, resist reflectance variation, background texture.

This separation means the base model can be updated (improved defect coverage, finer resolution) without invalidating your accumulated recipe calibrations — critical in production environments where re-qualification is expensive.

False-Positive Control Deep Dive
Optical Sensor Frame Input Base CNN Model Feature extraction — defect detection Updated per model version, not per recipe Per-Recipe Calibration Layer Tool-specific thresholds — auto-updates from lot data Defect Class + Confidence + Coordinates

Request a technical briefing

45-minute technical walkthrough with our engineering team. Covers model architecture, calibration methodology, and integration requirements specific to your process configuration.