New tools for applied materials use artificial intelligence to detect errors on chips

SAN FRANCISCO (Reuters) – A new semiconductor manufacturing technology from Applied Materials uses artificial intelligence, or AI, to detect errors in chips more effectively, the US company said Tuesday.

Modern chip manufacturing takes place in factories that cost up to $ 18 billion each and require hundreds of separate steps. Ensuring chips roll off the factory line with no errors in features that are only a few nanometers wide is critical to the ability of companies like Intel Corp, Taiwan Semiconductor Manufacturing Co, and Samsung Electronics Co Ltd to make a profit.

The new tools applied are meant to inspect those chips at various points during the manufacturing process. A new optical scanner, essentially an extremely advanced camera that Applied calls Enlight, quickly scans a silicon wafer for problem areas for about 15 minutes, then an electron microscope zooms in for a closer look.

The problem Applied aimed to solve with AI is that electron microscopes are accurate but slow. An initial optical scan could find a million possible problem areas on a silicon wafer and it would take electron microscopy days to examine each of those areas – and much of that time would be wasted, because only a fraction of the problem areas are those that chip industry veterans call it “killer” defects that would cause the chip to malfunction.

The new artificial intelligence technology, which Applied calls ExtractAI, only needs to check about 1,000 of those possible trouble spots with the electron microscope to predict where the biggest problems will be. Keith Wells, group vice president and general manager for imaging and process control at Applied, said AI-based control only takes about an hour.

“It is economical for the customer to do it on every wafer,” Wells said in an interview. “We are confidently telling you that these are the truly deadly flaws.”

Applied has been testing the system with customers since last year and has claimed to have generated over $ 400 million in revenue so far.

Reporting by Stephen Nellis; Editing by Rosalba O’Brien