Fujitsu’s e-beam offers fast, inexpensive prototyping path for 65 nm designs

The problem of getting affordable prototype dice in advanced process nodes is defined in part by the time and cost of producing mask sets. Fujitsu Ltd.’s E-Shuttle operation in Mie, Japan has been using a direct-write e-beam system to circumvent mask-making for the most critical layers of 65 nm prototyping runs. In a paper at SPIE this week, the company described a series of developments to improve the throughput and quality of results in this process.

In this flow, Fujitsu takes GDS-II design data from a customer, and separates the data for the first roughly dozen masks from the rest of the file. The higher layers go to a conventional mask-making process. But the critical first 12 layers are printed without photomasks—by writing directly onto the resist with an e-beam system.

The e-beam process is far from simple. To begin with, it’s slow: Fujitsu’s target, which has demanded considerable engineering, is 0.5 wafers/hour throughput. Just getting this far has required a careful trade-off between writing speed and resist sensitivity. The trade-off is necessary because increased resist sensitivity comes with increased line-edge roughness, which quickly becomes an unacceptable source of electrical variations in the finished circuits.

Beyond raw speed, Fujitsu developers found three main problems to be overcome: resist collapse, variation in critical dimensions due to instability in the variable-shaped beam used in the e-beam system, and interactions between the e-beam and conductive layers lying below the resist. Takashi Maruyama, manager of the EB Lithography Department at Fujitsu’s Mie facility, described how each of these had been addressed.

Resist collapse is a mechanical problem. At 65 nm, features can be so small and closely packed that after developing and cleaning, there are tall, narrow, delicate features in the remaining resist, possibly with heights several times their width. A cross-section photo of the resist can look a great deal like a side-on view of a comb. These structures, which after all are just polymer material, are subject to simply falling over and sticking to each other, ruining the pattern.

Maruyama said that the developers found a way of tuning the base polymer material to strengthen it. And by developing a process that simply rounds off the tops of narrow resist structures to reduce surface attraction forces between features, the problem was brought within acceptable production limits.

Beam instability required a more ambitious solution. Normally, a variable-shaped beam e-beam system uses a pair of rectangular apertures. By moving these apertures back and forth with respect to each other, the system can produce one arbitrarily-shaped rectangular pattern per shot.

But the Fujitsu team identified problems with the stability of the beam in this approach. So they created a block mask in the e-beam system that contains, instead of a simple rectangular aperture, a “character set” of specific patterns. Shaping the beam by passing it through one of these patterns substantially improves beam stability compared to passing the beam through a large aperture, leading to better CD control.

The block mask only allows the printing of a “character set” of pre-defined shapes. But with 12 areas available on the block mask, and room for about 100 “characters” in each area, the system has a more than adequate vocabulary of about 1200 shapes. In addition, according to Maruyama, the team devised a way to define complex, multi-feature characters on the mask and then partially block them off, allowing a single shot of the e-beam to write multiple features onto the wafer. This improved system throughput without sacrificing the gain in beam stability. The team’s studies indicate that using this partial-blocking approach can reduce the number of shots on a layer by as much as a factor of 20, and by about a factor of 4 on the average—a substantial boost to throughput.

Finally, the team studied the interactions that occur between buried layers and the e-beam on the surface of the wafer. They discovered that metal layers and vias beneath the surface where the beam is striking the wafer cause back-scattering of electrons, up to a depth of at least 12 layers. This back-scattering increases the effective dose in areas above the buried metal, distorting the resulting feature. The solution here was computation—Fujitsu developed an algorithm, executing on a cluster of PCs, that reduces the dose based on the number of underlying metal structures for each shot in the exposure process.

 The end result, according to Maruyama, is that Fujitsu has achieved their throughput, CD-stability, dose stability, and resist-strength targets, and is now accepting 65 nm designs for prototype production with the system. Now the team is looking on to 45 nm. The resolution and stability of the system don’t appear to be challenges, he said. But there is one big issue: resist at the 45 nm node has less than half the sensitivity to e-beam energy as does the 65 nm resist. Possible solutions include further reduction in shot counts to improve throughput at longer doses, increasing the beam current, or advances in resist technology—within the limits imposed by line-edge roughness. All these are possible, but at the moment, Maruyama worries, there is no clear roadmap.