Graph Attention Networks Boost Virtual Metrology for Chip Production

Leveraging Graph Attention to Map Process Interdependencies

Researchers from Arizona State University and Intel’s Foundry division unveiled a new graph‑attention model for virtual metrology on June 2, 2026. The study, presented at the International Conference on Machine Learning for Manufacturing, demonstrates how AI can predict wafer‑level measurements without costly physical probes.

The team combined graph neural networks with attention mechanisms to capture complex relationships among process sensors, equipment, and material flows. By training on historic fab data, the model forecasts critical dimensions such as line‑width and film thickness with sub‑nanometer accuracy. This approach reduces the need for invasive test structures, cutting both time and material waste.

The researchers also introduced a dynamic updating scheme. As new sensor data streams in, the graph recalibrates its attention scores, keeping predictions aligned with shifting process conditions. This adaptability proved crucial during a scheduled equipment upgrade, where the model maintained stable accuracy despite altered thermal profiles.

Can the Technique Scale to Next‑Generation Nodes?

Scaling the approach to 3‑nm and beyond poses challenges, chiefly the explosion of sensor nodes and the need for faster inference. The authors propose hierarchical graph constructions that group related sensors, reducing computational load while preserving critical detail. Preliminary simulations suggest the method can handle ten‑fold larger graphs with only modest latency increases. Intel’s fab chief, Carlos Mendes, noted that „if the model can keep pace with our production cadence, it will become a core component of our digital twins.”

The breakthrough signals a shift toward fully data‑driven fabs, where AI predicts outcomes before silicon ever leaves the cleanroom. By minimizing reliance on physical metrology, manufacturers can accelerate cycle times, lower defect rates, and improve yield forecasts. Continued collaboration between academia and industry will be essential to refine the models and integrate them into existing fab control systems.

Frequently Asked Questions

What is virtual metrology? Virtual metrology uses software models to estimate physical measurements of wafers, eliminating the need for direct probing during production.

How does graph attention improve prediction accuracy? It assigns higher importance to sensor connections that most influence a target measurement, allowing the network to focus on relevant data while ignoring noise.

Will this replace traditional metrology entirely? Not immediately. The technology complements existing methods, offering early warnings and reducing probe usage, but physical verification will remain a safety net for critical steps.