SSAR: Surface Anomaly Recognition Model

This repository contains the final evaluation results for the SSAR (Semiconductor Surface Anomaly Recognition) model. The project focuses on high-precision defect detection and classification in semiconductor manufacturing.

Performance Overview

The model achieves near-perfect classification across 9 distinct categories. The evaluation was performed on a balanced test set of approximately 3,600 samples.

Final Confusion Matrix

Below is the visualization of the model's performance on the test dataset:

Analysis of Results

The model demonstrates exceptional accuracy, particularly in identifying critical electrical defects like Shorts, Opens, and Corrosion.

Key Observations:

• High Precision: Categories such as Corrosion and Shorts show 0% false positives from other classes.
• Minor Confusion: There is a slight overlap between Clean and Other categories (approx. 4% error rate). This is expected as "Other" often contains subtle artifacts that resemble a clean surface.
• Particle Sensitivity: The Particles class acts as a slight attractor for minor misclassifications from LER and Scratches, likely due to shared geometric features.

Classes Identified

Category	Description
Clean	No defects present.
Corrosion	Chemical degradation of the metallic surface.
LER	Line Edge Roughness; deviations in the edge of a feature.
MalformedVia	Incorrectly formed vertical interconnect access points.
Opens	Breaks in the intended conductive path.
Particles	Foreign material or dust on the surface.
Scratches	Physical abrasions on the wafer surface.
Shorts	Unintended connections between conductive paths.
Other	Miscellaneous anomalies not covered by the above.

---