Using Graph Attention for Virtual Metrology in Semiconductor Manufacturing (Intel Foundry, ASU)
Original reporting by Semiconductor Engineering
The precision required for modern semiconductor manufacturing is staggering. Operating at nanometer and angstrom scales, the quality of every film layer deposition is paramount, demanding real-time, accurate measurements. Yet, traditional physical metrology faces significant hurdles: it’s slow, costly, and cannot sample every wafer, creating bottlenecks in high-volume production. This has led the industry to embrace virtual metrology (VM), where artificial intelligence predicts wafer characteristics directly from equipment sensor data. While effective, many existing VM models are largely correlation-driven, struggling to capture the complex, structured dependencies among diverse process variables and often lacking transparency in their predictions.
A novel approach
Addressing these limitations, researchers from Arizona State University and Intel Foundry have unveiled a sophisticated graph attention-based VM framework. Published in a recent paper, their innovation integrates temporal feature learning—extracting insights from high-frequency sensor traces—with a unique parameter-to-layer graph attention mechanism. This architecture allows the model to understand directional dependencies, enabling each film layer to intelligently aggregate relevant process information. Tested against industrial deposition data, the framework not only significantly improves prediction accuracy for film thickness compared to current benchmarks but also offers crucial interpretability. By analyzing learned attention weights, engineers can now discern physically consistent relationships between process parameters and film layers, providing invaluable insight for process monitoring and optimization.
The joint research from Arizona State University and Intel Foundry marks a significant advance in semiconductor manufacturing with their graph attention-based virtual metrology (VM) framework. This innovation directly tackles the long-standing challenges of physical metrology—measurement latency, cost, and sampling constraints—by offering superior predictive performance and, crucially, interpretable insights into complex film deposition processes. By leveraging AI to model intricate parameter-layer dependencies and temporal features, even at nanometer and angstrom scales, the system provides real-time, high-precision feedback essential for robust process control. This capability ensures proactive identification and mitigation of manufacturing issues, enhancing efficiency, reducing costly defects, and accelerating yield improvements in high-volume production environments.
Future Trajectory
This breakthrough extends far beyond the immediate benefits of faster, more reliable chip manufacturing. It represents a pivotal step in the evolution of AI’s role in industrial processes, shifting from purely correlation-driven models to intelligent systems that deeply understand and explain physical process behavior. Such interpretable AI is vital for accelerating innovation cycles, enabling engineers to refine designs and processes with unprecedented clarity and confidence. The success of this framework in the highly demanding semiconductor sector offers a powerful blueprint for other high-precision manufacturing industries facing similar metrology bottlenecks, from advanced materials to biomedical devices. Ultimately, this research underpins the broader drive towards increasingly automated, resilient, and intelligent manufacturing ecosystems, fortifying global supply chains and powering the next generation of technological progress across diverse sectors.