The last few generations of ICs incorporate incredibly small, physically complicated structures that are exceedingly difficult to manufacture, and now production processes have become so complex that inspection equipment cannot keep up. Applied Materials is remedying that by pairing its new Enlight optical inspection system with its SEMVision e-beam monitoring tool, and by supplementing the combination with a new AI-based system called ExtractAI that helps direct the inspection process.
Production yield has always been a critical factor in semiconductor manufacturing, and that’s why IC manufacturers have always monitored wafers as often as they practically could throughout the production process. But producing integrated circuits has become so complex, the legacy approach to wafer inspection is no longer viable, according to Rafael (Rafi) Benami, vice president of Applied Materials’ process diagnostics and control products group.
IC structures — and the defects that can kill ICs — are now so infinitesimally small that they’re nearing not only the physical limits of silicon, but also nearing the resolution limits of optical inspection systems. That’s a problem because having increasingly smaller features means that the most miniscule mis-formations and ever-tinier particles create problems. E-beam inspection systems, meanwhile, have superior resolution to optical systems, and are therefore more accurate, but they’re significantly slower.
Production in the most recent nodes relies on multi-patterning (which is to say: far more steps), which translates into more points of possible failure, more possibilities for contamination, and therefore more potential inspection points. It is possible to increase the number of inspection points, but the cost goes up in a linear way, Benami said.
As a practical matter, something has to give, and what gives is often inspection.
That’s dangerous. Anything that could lead to lower yields is bad in several different ways. The longer a problem goes undiscovered, the more expensive it gets to fix. The cost of finding and a fixing a problem during R&D is bad enough, but the cost of finding and fixing a problem once into production can cost dearly; with a logic device at the 3nm node, a single week of downtime translates into $25 million in unamortized depreciation, Benami said. Similarly, a week of downtime producing DRAM costs 2% of annual revenue, plus price erosion.
Applied’s response is to combine an improved optical system with an e-beam system, and use the combination in as efficient a way as possible, hence the artificial intelligence (AI).
First, the company is pushing the speed limits and the resolution of optical systems with its new Enlight product. The system has both lightfield and greyfield (the common industry term is “darkfield”) detectors; the former collects light directly reflected from flat surfaces, while the latter picks up light that scatters from angled surfaces. The new system can collect more yield-critical data per scan, Applied said, “resulting in a 3x reduction in the cost of capturing critical defects as compared to competing approaches.”
This allows chipmakers to insert many more inspection points in the process flow, according to the company. The greater volume of data improves the system’s ability to predict yield excursions before they occur, and enables root-cause traceback to accelerate corrective actions, the company added.
Applied pairs its new Enlight optical system with the latest version of its SEMVision e-beam inspection system, the G7.
Getting the two to work together well is one of the key places where the new ExtractAI technology comes in.
The company explained that high-end optical scanners generate millions of nuisance signals — noise. Sifting out actual defects from the noise is an ongoing problem. Applied says that its AI can learn to classify “specific yield signals so that by inference, the Enlight system resolves all of the signals on the wafer map, differentiating yield killers from noise. ExtractAI technology is incredibly efficient; it characterizes all of the potential defects on the wafer map after reviewing only 0.001% of the samples. The result is an actionable map of classified defects that accelerates semiconductor node development, ramp and yield.”
The system relies on the fast optical component for first-pass scans. It analyzes the data in real-time, and then directs the e-beam system to more precisely scrutinize those sites most likely to be actual defects, Benami explained.
The company said its installed base of SEMVision G7 systems is already compatible with the new Enlight system and ExtractAI technology. Applied says it has been developing Enlight and ExtractAI since 2016, and that the new inspection system is in use in dozens of fabs all around the world.
Asked about the price, Benami responded: “It costs a lot of money. Relative to our competitor, though, it’s much more economical due to the speed and sensitivity.”