Air gaps may be the ultimate low-k dielectric—but there are challenges
The search for low-k material to put between the interconnect lines on an IC has led us from oxides of silicon to strange low-k materials, to even stranger porous materials. The obvious end-point of this search is nothing: the best possible low-k dielectric material is air, or vacuum. And in fact research is actively under way at a number of labs to use air gaps in between metal segments in the interconnect stack.
The idea isn’t new. TriQuint Semiconductor used air gaps maybe 20 years ago to fight parasitic capacitance in their GaAs circuits. But it has come to the fore recently as process engineers have seen the end of the roadmap approaching for low-k dielectric materials. Once you have devised a gel that is almost entirely bubbles, and it’s still not good enough, there’s nothing left to do but eliminate the gel and keep only the bubbles.
But actually fabricating interconnect stacks with air gaps between the conductors has proved to be a very non-trivial, and so far, non-manufacturable, idea. Papers at the International Interconnect Technology Conference this week examined the state of air gap technology, showed remarkable progress, and suggested some ways forward.
A paper by Satya Nitta et al. of IBM TJ Watson Research Center quickly surveyed the current approaches to creating air gaps, and then described IBM’s favored approach in some detail. Basically, according to Nitta, there are two ways to make an air gap. You can do a relatively conventional damascene process, but with a sacrificial material used as the inter-metal and inter-layer dielectrics, and a strong cap over each layer. Then you can cook out the sacrificial material, leaving—ideally—empty space between, above, and below the interconnect lines.
Potentially this approach should be very rewarding, because it could provide an air barrier around nearly all of any critical metal link, very substantially reducing parasitic capacitance and inter-metal coupling. But the problems are also formidable. Among the obvious ones are design constraints: you have to leave enough material in place to mechanically support the metal links, or they will collapse into each other. And you have to leave enough holes in the cap so the sacrificial material, once it has been vaporized, can escape.
There are more detailed problems, according to Nitta. One is that the polymer materials that are willing to sacrifice themselves do not integrate easily into existing processes. Cap materials are critical to the success of the process, because the cap can buckle or crack under external stress if it is not very carefully designed. So extensive use of dummy metal is indicated. Nitta also warned that the sacrificial approach would require a redesigned back end of line (BEOL)—a concern that later papers did not appear to support.
Not surprisingly after all these observations, IBM’s approach is quite different. The company has decided to form air gaps not by use of a sacrificial material, but by conventional etch-back techniques. And they have chosen to only form air gaps in between the metal segments in a layer—not between layers. So the metal is always supported from below at least by the inter-layer dielectric, and often by underlying metal as well. This greatly improves mechanical strength. But just to be sure, IBM employed finite element modeling of the multi-layer structures to come up with some design rules to address structural strength. To avoid the widely-discussed problem of dealing with vias in the air-gap structure, IBM simply decided to put exclusion zones around the vias, and not try to form air gaps there.
This approach, Nitta said, does not require changes to the existing low-k BEOL. It does require, in the way IBM is doing it, an additional etch-back mask that controls the selective removal of the low-k material. The paper also described a self-assembling etch-back technique, but that does not appear to be the company’s preference.
But with all these simplifications, does it work? Nitta presented measured data suggesting that IBM’s approach achieved a reduction in the effective k value from 3 to 2.1—which translates into about a 35 percent reduction in capacitance.
Meanwhile the sacrificial-material approach had its sponsors as well at the conference. A paper from Toshiba Corp. proposed reducing the cost and process complexity of an air gap approach by employing a sacrificial material on multiple layers, leaving the material in place until all the layers have been fabricated. Then the proposal is to drill vent holes through the interconnect into all the layers where the material must be removed, and then to remove all the material at once. The paper also discussed ways of dealing with moisture formation in the air gaps—a common side effect of processing—and oxidation of the metal segments.
A second paper, from ST Microelectronics Crolles and two research organizations in Grenoble, France, reported on a successful application of sacrificial-material air-gap design to a 65 nm process running 300 mm wafers. The approach showed an extracted effective k of 2.2, using an ordinary silicon-dioxide dielectric material. The authors estimated that if their air-gap approach were used with an aggressive low-k material with a k=2.5, the effective result would be a k below 2.0. The paper also pointed out that while the research had shown manufacturability on a 65 nm process, the BEOL was little changed, and the air gap technique could be used in older processes to selectively reduce the capacitance on critical paths.
Taken together, the papers seem to indicate not only that there is considerable promise in using air gaps to isolate critical metal segments, but that the industry is moving toward manufacturable approaches to doing just this. Nitta went so far as to say that IBM will apply their air-gap technology to their full low-k roadmap in the future. Expect the foundries to follow, but not quite so quickly. This will give designers one more weapon against interconnect capacitance, but at the cost of one more increment of design complexity.