Today’s chip manufacturing is a high-stakes process where even minor errors can lead to significant losses. With intricate workflows involving thousands of precision steps, the industry has long struggled to identify defects early and optimize production.
Now, IBM Research is harnessing artificial intelligence to transform defect detection, process understanding, and workflow management, setting the stage for a more efficient and reliable future in chip fabrication.
Spotting Defects at the Source
Traditional approaches to tracking defects in silicon wafers are manual, time-consuming, and often miss the exact cause of failures.
IBM’s AI-powered systems, including platforms like SiView and Intelligent Fab, analyze massive amounts of process data to pinpoint when and where defects emerge.
These algorithms assign responsibility scores using an advanced technique called the Trajectory Shapley Value, which reveals critical moments in the manufacturing process.
This enables manufacturers to catch issues early and stop the production of defective wafers, conserving valuable time and resources.
Understanding Process Interdependencies
Beyond individual defects, the production of semiconductors involves complex interactions between various manufacturing steps.
IBM’s innovative “proc2vec” algorithm, inspired by language processing technology, maps out hidden relationships among process stages.
This method improves predictions by considering how different steps influence each other, much like understanding context in human language.
- Proc2vec uncovers intricate process relationships
- Enhances accuracy in defect prediction with historical data
- Identifies anomalies tied to factors like extended wait times or process delays
Rethinking Workflow Management
Workflow congestion, or WIP (work-in-progress) bubbles, has traditionally been modeled with basic queuing theory. However, these models often fall short in explaining the unpredictable movement of lots through a factory.
IBM researchers have applied the Hawkes process, a statistical model that considers historical events, to better predict real-world delays. Their findings reveal that cycle times can be much longer than previously estimated, especially under high tool utilization or fluctuating lot arrivals.
This challenges industry norms and calls for more sophisticated analysis tools to reflect the complexities of modern chip fabs.
The Road Ahead: Merging AI with Physics
Although these AI-powered solutions are still evolving, IBM’s team sees even greater potential by integrating deeper, physics-based insights into their models.
By deploying advanced analytics directly onto production lines, the aim is to reduce defects and boost yield across the semiconductor sector.
Artificial intelligence is pushing semiconductor production into a new era, not just by predicting defects, but by fostering smarter, more proactive manufacturing strategies.
As AI matures, it will become a core driver in optimizing reliability, efficiency, and scale in chip manufacturing.
Source: IBM Research Blog
AI is Revolutionizing Chip Manufacturing from Within