Revolutionizing Semiconductor Inspection with AI: The Two-Stage Approach

Detecting minuscule defects in semiconductor lithography is no small feat. Essential for quality control, it involves identifying defects like bridges, burrs, pinches, and contamination. A new two-stage vision-language framework is set to revolutionize how we approach this challenge.

Initial Detection and Its Challenges

In the first stage, a model named Qwen3-VL is fine-tuned using a method called LoRA, serving as a vision-language adapter. This setup predicts defect counts, defect categories, and generates normalized bounding boxes from lithography images. The paper's key contribution: enhancing initial detection capabilities without the usual manual inspection load.

However, relying solely on fine-tuning isn't without pitfalls. Common test-time errors such as false positives, missed defects, and misclassification still occur. These inaccuracies can hinder productivity and raise costs, driving the need for refinement.

The Role of Refinement

The second stage comes into play by introducing a refinement module. This module is trained on the mistakes made by the first-stage predictions, along with their corrected labels. By learning from these initial failures, the refinement process elevates the system's defect inference beyond standard single-stage fine-tuning.

But how effective is this two-stage approach? The ablation study reveals that incorporating refinement significantly reduces error rates. It demonstrates that continuous learning from past errors is essential to improving AI model performance.

Implications for the Semiconductor Industry

Why should this matter to the semiconductor industry? It's simple. Enhanced inspection accuracy leads to fewer production halts and better product quality. This directly impacts profit margins and market competitiveness. With semiconductor production being a cornerstone of modern technology, any advancement in this field can have wide-ranging implications.

However, one must wonder, is this refinement process sustainable in the long term? While it shows promise, the constant evolution of defects and patterns may require ongoing adjustments to the model, potentially affecting scalability and cost-effectiveness.

Code and data are available at the associated repository, ensuring that this research is reproducible and open for further exploration. This builds on prior work from the field, pushing the boundaries of what's possible with AI in industrial applications.