Rad AI Demonstrates Wafer Defect Detection Benchmark on Full WM-811K Dataset

According to a recent LinkedIn post from Rad AI, the company replicated an IEEE-published study on semiconductor wafer defect detection and challenged its experimental design. The post asserts that the original authors reported 91.7% accuracy using a CNN on the WM-811K dataset but relied on only 902 carefully selected images, representing about 0.5% of available labeled samples.

Rad AI’s post indicates that it reran the task on the full dataset of more than 172,000 labeled wafer maps, using a ResNet architecture and preserving natural class imbalance while training on CPU hardware. The experiment reportedly achieved a validation accuracy of 95.25% and an F1 score of 95.28% after 76 epochs, with a training time of roughly five hours under what the company describes as resource-constrained conditions.

The post emphasizes themes of research integrity, transparency, and realistic handling of class imbalance, noting that its code is open and methodology described as fully documented. For investors, this positioning may suggest internal technical competence in applied deep learning, particularly in semiconductor defect inspection, and a potential edge in quality-sensitive automation markets where dataset rigor and reproducibility are becoming competitive differentiators.

If Rad AI can translate these benchmarking results into production-grade inspection solutions, it could enhance its appeal to semiconductor manufacturers seeking yield improvement and more reliable defect detection workflows. More broadly, the focus on open methodology could support trust-building with industrial partners and may help the company participate in a wider ecosystem of research collaborations and validation efforts in machine learning for manufacturing quality control.