Rad AI Demonstrates Full-Dataset Wafer Defect Detection Performance

According to a recent LinkedIn post from Rad AI, the company replicated an IEEE-published study on semiconductor wafer defect detection and questioned aspects of the original methodology. The post notes that while the original authors reported 91.7% accuracy using a CNN on the WM-811K dataset, they apparently relied on 902 balanced images, representing about 0.5% of more than 172,000 labeled samples.

Rad AI’s post indicates that it reran the experiment on the full dataset using a ResNet architecture, preserving natural class imbalance and training on CPU-only resources. The company reports achieving 95.25% validation accuracy and a 95.28% F1 score after 76 epochs, with training time of about five hours, and emphasizes open code and transparent methodology as key elements of its approach.

For investors, the experiment suggests Rad AI is positioning its technology and team as capable of handling large, realistic semiconductor defect datasets under resource constraints. This may strengthen the company’s credibility in high-value industrial automation and quality-control markets, where robust model performance, reproducible research, and data integrity are critical to adoption and potential commercial partnerships.

The post also underscores issues of class imbalance, dataset utilization, and research integrity in applied machine learning, which could resonate with enterprise customers concerned about reliability and auditability. If Rad AI can translate these technical capabilities into deployable inspection solutions, it may improve its competitive standing in semiconductor manufacturing workflows and related AI-driven inspection applications.