The stuff of tomorrow's computers is taking shape in IBM's laboratories.
In a world increasingly defined by the flow of data, it is easy to forget a fundamental fact about
bits and bytes: ultimately they must reside in chunks of matter. Data must be embodied in marks on
a piece of paper, a swarm of electrons in an integrated circuit, a herd of magnetic spins on the
surface of a hard disk or patterns of light in a computer display.
The ability to find
materials that can be processed, combined and formed
into ever smaller, faster and more reliable devices has largely defined the progress of computer
technology over the past
half century. Today, the limits to such progress are looming.
The materials that have served the industry so well for so long will need to be reworked,
rethought and, in some cases, replaced. In IBM's research laboratories, numerous projects
are devoted to "frontier materials" -- that is, novel substances that can do a job better
than existing ones or that can do something no other material can.
Some of the research, says Thomas Theis, director of physical sciences at IBM's Thomas J. Watson Research Center, is aimed simply at extending current technologies -- giving engineers the materials they need to make incremental improvements in existing devices. Other research is aimed at opening up completely new capabilities, as happened, for example, when IBM Fellow Stuart Parkin synthesized materials that exhibit giant magnetoresistance (large changes in resistance in response to an applied magnetic field) -- with low magnetic fields and at room temperature. Parkin's work allowed the development of a fundamentally different type of read head for magnetic hard disks, one that is far more sensitive than conventional read heads. Indeed, research in that area continues to yield a stream of innovations, inventions, and discoveries.
Both sorts of investigation -- incremental and revolutionary -- are necessary, Theis says, since incremental advances tend to bring a technology to the point where it can no longer be easily improved and where a new approach is needed to continue the progress. Moreover, what was speculative exploration of new materials five years ago may be incremental materials engineering today.
The frontier materials under investigation at IBM Research range from new takes on conventional stuff -- ceramics,
polymers and alloys -- to exotic materials like microscopic carbon tubes. But all are being considered for innovative uses.
Those that meet the stringent requirements of product development and manufacturing will be the materials upon which the
computer industry of the future will be built.
THE SHRINKING INSULATOR
The success of IBM's low-resistance copper technology introduced last year highlights the fact that integrated-circuit chips consist of more than just the semiconductors, such as silicon, that most people associate with them. The chips also require conducting materials, such as aluminum and now copper, to carry electric current around the chip, and insulators, which block current flow.
The insulators on a chip are particularly important materials, for they perform several diverse functions. Insulators are used to create the memory cells that store data on chips. In a field-effect transistor -- the basic component of computer chips -- a layer of insulator isolates the gate from the channel that connects the source and the drain. And insulators are used to separate the layers of wiring on a chip, preventing current from leaking from one circuit to another.
In today's most widely used chip technology -- CMOS (complementary metal-oxide semiconductor) -- the insulator is silicon dioxide. It is well suited for its job not only because it is a good insulator but because it is compatible with silicon and can easily be "grown" where needed during the fabrication of a chip. Furthermore, the fabrication process has been optimized for silicon dioxide, and any replacement will naturally force a reworking of that process. But in spite of the central role the material has played in the semiconductor industry, the consensus is that replacements must be found.
This is the case, for instance, with the memory cell on a dynamic random access memory (DRAM) chip. The cell includes a capacitor -- a pair of conductors separated by a silicon dioxide insulator -- which is used to store a tiny electric charge; the presence or absence of a charge in the cell corresponds to a 1 or a 0 in the memory. As Jean-Pierre Locquet of IBM's Zurich Research Laboratory explains, each cell must be able to store a minimum amount of charge to function effectively. But to squeeze more memory onto a chip, the cells have had to shrink, while continuing to hold the same amount of charge. To date, designers have compensated by modifying the shape of the cells, making them narrow and deep to maximize their charge-storage capacity. But the diminishing size of cells makes this more and more difficult. "At some point," Locquet says, "we will have to switch from silicon dioxide to a material with a higher dielectric constant" -- that is, a material that can naturally store more charge in a given volum
e.
Locquet and his colleague Jean Fompeyrine are looking at perovskites, a family of minerals either found in the earth's crust or created synthetically, which have some of the highest dielectric constants known. Unfortunately, unlike silicon dioxide, the dielectric constants of perovskites vary with their form. A solid crystal of barium titanate, for example, can have a dielectric constant between 5,000 and 10,000, compared with a dielectric constant of 4 for silicon dioxide, but when barium titanate is made into a thin film -- the form it would take in a computer chip -- its dielectric constant drops precipitously. A typical film that is 20 nanometers (billionths of a meter) thick has a dielectric constant of only about 200, higher than silicon dioxide but disappointing relative to bulk barium titanate. Furthermore, since barium titanate reacts strongly with silicon during processing, it cannot easily be formed into the three-dimensional trench structures used to increase the storage capacity of silicon dioxid
e memory cells. To turn barium titanate into a useful material for memory cells, the researchers will have to find some way to substantially increase the dielectric constant in thin films of the stuff.
To that end, Locquet and Fompeyrine are modifying the atomic spacing in the thin films by growing them layer by layer on a substrate whose own atomic spacing is slightly larger or smaller than that of the perovskite. As the spacing of the perovskite film tries to align with that of the substrate, it is stretched or squeezed. The researchers have shown in other materials, such as perovskite superconductors, that such manipulation of the atomic spacing can have far-reaching effects, such as transforming a metal into an insulator. Now they hope to control the dielectric constant of barium titanate in a similar manner. They're halfway there. "So far, we've mainly succeeded in reducing the dielectric constant," Locquet says with a laugh. "But thanks to a detailed microstructural analysis, we are starting to understand why." Assuming the researchers can learn to increase the dielectric constant as well, barium titanate could be a promising material to replace silicon dioxide in memory cells.
The problem with insulators in CMOS technology affects more than just the capacitor in memory cells. Beneath the gate of every CMOS transistor is a thin insulating layer of silicon dioxide. "As we have scaled down our CMOS devices further and further," explains Susan Cohen, manager of advanced gate dielectrics at Watson, "the silicon dioxide layer has had to become thinner and thinner." But as that happens, the layer -- typically 3 to 4 nanometers thick in current manufacturing -- loses its insulating capacity, and the amount of current that leaks through the insulator grows exponentially. "We are approaching the limit, thought to be somewhere below approximately 2.0 nanometers, where silicon dioxide will no longer be useful as a gate insulator," Cohen says.
Thus researchers at Watson are exploring alternatives. To work, they should be compatible with the existing CMOS technology -- in particular, they must not react too strongly with silicon at the high temperatures used in processing the chips -- and they must have a higher dielectric constant. "The higher the dielectric constant," Cohen explains, "the thicker the film can be, and the lower the leakage."
The candidates include aluminum oxide, zirconium oxide, yttrium oxide and silicon dioxide mixed with transition or rare-earth metals such as zirconium or lanthanum. IBM scientists are testing the materials' electrical properties, including their dielectric constants, as well as their physical and chemical characteristics, such as how they interact with silicon. To complicate matters, all these properties depend in part on how the materials are processed -- how they are deposited on the chip and how succeeding layers are deposited atop them. But if chips are to improve at their current pace, Cohen says, silicon dioxide will have to be replaced as the gate insulator within six years.
At IBM's Almaden Research Center, meanwhile, Bob Miller and his group are wrestling with a very different problem, but one that also has its roots in the steady shrinkage of features on a computer chip. As the wiring on the chip becomes smaller and more closely packed, the speed at which an electrical signal passes through the wires is limited by an effect called capacitive coupling. The phenomenon is similar to what happens when an electric charge is stored in a memory cell: the silicon dioxide serving as an insulator between two layers of wiring accumulates a charge when currents pass through the wires. It takes time for this unwanted charge to fully build up, and, until it does, the signal is not transferred between transistors on the chip. The delay is proportional to how much charge the insulator will hold, which in turn depends upon its dielectric constant. And because the chips of 2003 or 2004 are expected to have 6 billion transistors and 10,000 meters of wiring, it will be important to squeeze out ev
ery possible picosecond of delay.
Miller is faced with the reverse of Locquet's and Cohen's quests: he wants to find a replacement material with a lower dielectric constant than silicon dioxide. Unlike Locquet and Cohen, however, he believes he has found a way to have his cake and eat it -- to change the dielectric constant but not have to move to a completely new material.
Miller's solution is, in essence, to make an insulator that resembles Swiss cheese. His group at Almaden has developed a simple way to create porous organosilicates -- films of an organic/inorganic hybrid material that is similar to silicon dioxide but is filled with tiny air-filled holes. The holes are created during processing by organic molecules that clump together into tiny domains that are later baked away, leaving air pockets in the organosilicate. By altering the size and number of the pores, Miller says, it is possible to engineer these materials to have a wide range of dielectric constants -- as low as 1.5 (compared with 1 for air).
The pores do create a number of problems, however. "Porosity," Miller says, "is good for only one thing -- a low dielectric constant -- and for everything else it is bad." The pores mechanically weaken the insulator, for instance, and they slow heat transfer, making it more difficult to dissipate the heat generated in the chips.
Still, the lure of a material whose dielectric constant could be continually lowered to match the shrinking feature sizes of chips is hard to resist, and IBM, in partnership with Dow Chemical, has an Advanced Technology Program grant from the National Institute of Standards and Technology to study porous dielectric insulators for use on chips. Already the group has developed one material (porous methyl silsesquioxane, or Dendriglass) that meets several key requirements: it adheres well to silicon, survives high processing temperatures, and can be polished. If it stands up under lithographic processing and possesses the right electrical properties after it has been incorporated into devices, it could become a strong candidate for a future generation of IBM CMOS computer chips.
MAKING IMPRESSIONS
The lithographic processing of chips is itself highly dependent on just the right choice of materials. Consider the photoresist used to create the patterns of transistors, wires and other structures on silicon chips. The photoresist -- whose key component is a polymer, a material consisting of long chains of repeating subunits -- is formulated to become soluble wherever it is exposed to light. To make each of two dozen patterned layers on a chip, a layer of photoresist must first be deposited over the entire surface of the silicon wafer. When the photoresist is exposed to a laser beam that has been aimed through an intricately patterned mask, those parts of the photoresist struck by the light are readily dissolved away, while the unexposed sections stay put. The pattern of photoresist that remains is then used as a shield, either to protect the underlying layers from a beam of ions that etches away any exposed material or to define where new materials are to be deposited. When the remaining photoresist is rem
oved with an organic solvent, the silicon wafer is left with the desired pattern formed on its surface.
The success of this process is critically dependent upon the material used for the photoresist, notes Donald Hofer, manager of the lithography materials group at Almaden. Every few years, as chip features shrink, it is necessary to find a new material.
The problem, Hofer explains, lies with the short wavelengths of light that are necessary to create the patterns on chips. As holes in masks get smaller, light diffracts more, blurring the image. So shorter wavelengths must be used to lessen the diffraction. The current generation of lithography uses ultraviolet light with a wavelength of 248 nanometers, which will eventually allow chip features as small as 150 nanometers--about one-thousandth the width of a human hair. The next generation of lithography will use 193-nanometer lasers to create significantly smaller features.
The challenge for materials research, Hofer says, is that "most materials absorb far too much light at these wavelengths to be useful." The photoresist must be transparent enough to allow a given wavelength of light to penetrate the layer completely and render it soluble even at the bottom of the layer for the pattern to be uniform from top to bottom. The photoresist cannot be too transparent, however, or the light will pass through with no effect. "With every generational change in wavelengths," Hofer says, "you have to develop a whole new set of materials."
Hofer's group at Almaden has spent several years developing a photoresist material that will work at 193 nanometers. They modified a polymer that had the proper transparency and was resistant to ion etching to create the other properties needed in a photoresist, such as being soluble enough to be easily deposited on the silicon wafer. Hofer and his colleagues formed a partnership with B.F. Goodrich to produce the polymers they had designed and spent the next two years testing them. "No one in the world had made these polymers before," Hofer notes. Although there are certain rules of thumb about what to expect with a particular type of modification, the team could never be sure how a particular polymer would perform until it had been tested.
The two years of searching paid off with a photoresist that is now being tested in a manufacturing pilot line. Computer chips made with the new photoresist should be in production by the end of 2001. By 2003, if all goes according to schedule, chips with features as small as 100 nanometers should become available, marking the limit of what can be done with the 193-nanometer photoresist. Looking ahead, Hofer's group has already begun the search for a suitable photoresist material for the next, 157-nanometer generation of lithography, which will be used to form features as small as 75 to 80 nanometers.
SUPPLE SCREENS
IBM's photoresist work is an example of extending the current state of the art -- the new materials are needed simply to continue improving an existing technology. But other work in organic materials could usher in technologies that are unlike anything else seen before, such as computer displays that are as light and flexible as an overhead transparency.
Flat-panel displays like those on laptop computers consist of a thin layer
of liquid-crystal material sandwiched between two supporting sheets of glass and controlled by a network of thin-film amorphous-silicon transistors to create images. Researchers working with organic materials believe they can create displays that are lighter, more durable and more energy efficient.
Organic light-emitting devices, or OLEDs, have been under development for years at
various places around the world, notes Paul Seidler, head of science and technology at Zurich,
and the first products incorporating OLEDs are now appearing on the market. OLEDs are exceptionally
bright, can be made to shine in any color, operate efficiently at low voltages, and can potentially
be made very inexpensively. Furthermore, it should be possible to put them on plastic substrates
and create displays that are thin, lightweight, and flexible -- and perhaps even able to be
rolled up or folded when not in use. A team composed of researchers at Zurich and Almaden
is trying to develop the organic materials that will make such OLEDs feasible
(see sidebar,"Computing a better LED").
At Watson, meanwhile, Dave Mitzi is working to develop organic-inorganic hybrids that he hopes will do the job even better than purely organic materials. Mitzi wants to combine the robustness and thermal stability of inorganic materials with the advantages of organic molecules, such as simple processing and the ability to transform electricity into light very efficiently. To do this, Mitzi begins with organic-inorganic perovskites, materials made up of alternating inorganic and organic layers. In a typical hybrid, Mitzi explains, the organic layer contains relatively simple molecules. His scheme is to replace them with organic luminescent dye molecules, which give off light when tickled by an electric current. The inorganic perovskite framework would supply the necessary structural and electronic properties.
To date, Mitzi has managed to grow single crystals of such hybrids using an organic dye molecule and, with postdoctoral associate Konstantinos Chondroudis, he has incorporated the material into an electroluminescent device. While the preliminary devices are not particularly efficient at emitting light, the hybrid materials are easy to process. If, as Mitzi expects, he can improve the luminescent efficiencies, the hybrids could be a good choice for a new generation of light-emitting devices.
But if those devices, whether organic or hybrid, are to lead to flexible displays, researchers must also develop organic transistors. Silicon transistors would be impractical, explains Christos Dimitrakopoulos at Watson, because the silicon used in thin-film transistors is grown at 350 degrees C (660 degrees F), "and every good transparent plastic melts below that temperature." Organic transistors, on the other hand, can be formed at room temperature.
It has proved difficult, however, to find the right mix of materials to create a practical organic transistor -- in particular, one in which a relatively small voltage can switch on or off a relatively large current. Until this year, the best organic thin-film transistor was one developed in 1996 by scientists at Pennsylvania State University. Its performance was comparable to transistors made from amorphous silicon, the type of transistor normally used to control displays on portable computers. But it demanded about 100 volts, more than the batteries of a portable computer can generate efficiently.
In February, Dimitrakopoulos and co-workers announced a way around such high voltages. The key was Dimitrakopoulos's realization that the organic semiconductors contain "traps" that capture the charge carriers (electrons or holes), thereby reducing the current between the transistor's source and drain. At higher voltages, enough carriers are generated to occupy all the traps and still provide an adequate current. The voltage needed to reach the required concentration of charge carriers decreases as the dielectric constant of the insulator increases. Therefore, by replacing the usual silicon dioxide in the transistor's gate insulator with barium zirconate titanate, which has a higher dielectric constant, the team was able to reach the desired performance with just 5 volts.
The one remaining hurdle is to learn how to make the organic transistors
with a process that is more suited to commercial production. When Dimitrakopoulos created his experimental organic transistor, it was with a vacuum process that is likely to be more expensive than fabrication processes that deposit the organic material from a solution. "We are now working to change the processing," he says.
STORING INFORMATION
Of all computer components, none has improved faster than the magnetic hard-disk drive. From 1991 to 1997, the storage density on hard disks -- the number of bits per square inch of surface area -- doubled every 18 months, and in the past two years the pace has quickened, with a doubling every 12 months or so. In just 15 years, PC hard drives have swelled from 20 megabytes to almost 40 gigabytes.
Much of the progress in magnetic recording has come from engineering: improving the heads that write and read data, for example, and flying them closer to the disk surface, where they can read and write smaller areas on the disk. But that evolution is coming to an end, says Hal Rosen, manager of advanced disk technology at Almaden. If progress is to continue, new magnetic materials will be needed.
Each bit on a hard disk is created by magnetizing a small section of the surface of the disk. Looked at under a microscope, such a section can be seen to consist of a thousand or so individual grains, each a crystal formed by atoms arranged in a closely packed pattern. Each grain behaves like a tiny bar magnet. At first the magnetic field of each grain points in a random direction, but when a section of the disk is magnetized, all these bar magnets are aligned to point in one direction. Magnetic fields pointing right might indicate a 0, say, and a field pointing left a 1.
To cram more 0s and 1s onto a hard disk, one must shrink the surface area covered by an individual bit. If the number of grains in a bit dropped much below its current value, noise would make it difficult to distinguish a 0 from a 1. "So we have to make smaller and smaller grains," says Rosen. And for several generations of magnetic disks, IBM and other companies have done just that, finding new ways to process the material -- usually a mixture consisting mainly of cobalt, platinum, and chromium -- to make the grains smaller and more isolated from one another.
But this course, too, has its limits. The grains are now approaching a size where the normal jiggling around of atoms at room temperature is enough to flip the magnetization of a grain spontaneously from one direction to the other. This phenomenon, known as the superparamagnetic effect, will make it physically impossible to store information in grains smaller than about 10 nanometers using the current magnetic material. With today's grain sizes of 12 to 14 nanometers, IBM has demonstrated magnetic disks that hold 20 gigabits -- 20 billion bits of data -- per square inch. And thanks to its development of new cobalt-platinum-chromium alloys, IBM recently demonstrated a world-record 35 billion bits per square inch (see "A giant leap for magnetic media," page 3). With further materials engineering and improved signal processing, Rosen predicts a storage density of 100 gigabits per square inch will be reachable, making possible a 1 terabyte -- 1 trillion byte -- hard drive.
Beyond that, significantly different technologies will be needed. "There are a number of possible approaches," Rosen says, "all of which we're pursuing with vigor." To evaluate different approaches, IBM researchers are using computer simulations that take into account every part of the magnetic disk system, from the write and read heads to the magnetic recording material and the magnetic interactions among the grains -- all of which affect the ultimate performance of the disk.
One approach under investigation is perpendicular recording, in which the tiny bar magnets of the grains point up or down -- that is, out of or into the disk -- instead of right or left. Perpendicular recording uses thicker magnetic recording layers. Because a grain's resistance to flipping is proportional to its volume, bits with smaller surface areas would then be feasible. It would also be possible to work with hard-to-magnetize materials that, once magnetized, better resist superparamagnetic flipping. But taking advantage of these properties will demand the development of special read-write systems and materials optimized for perpendicular recording.
Rosen also foresees replacing the magnetic recording layer on the disk with a completely new type of material, such as an iron-platinum alloy or a cobalt-iron-oxygen alloy, whose resistance to spontaneous flipping is inherently much higher than the alloys now in use. Because they are so different from today's materials, however, they will demand a great deal of development before they are ready for market.
An even more exotic possibility is a "continuous exchange-coupled material," a type of material in which the grains interact magnetically, in effect creating a single large grain. If it could be made to work, such a material could have a much higher barrier to spontaneous flipping, allowing much smaller bit sizes.
Complicating all these possibilities is the fact that, on the tiniest scales, a material's behavior depends on its size -- a 5-nanometer grain of platinum, for example, has different properties from a 10-nanometer grain because the behavior of electrons is influenced by how tightly they are confined. At Watson, Christopher Murray and Shouheng Sun are working to help designers understand what to expect from magnetic materials at length scales so small they are years away from commercial production.
Murray and Sun study materials at lengths of 1 to 20 nanometers. "There are no current methods for manipulating matter in this regime," Murray notes. Traditional chemistry works at smaller scales, and engineering approaches, such as lithography, work at larger scales, so he and Sun have had to develop their own methods. To get nanometer-scale particles, they grow tiny crystals in solution, carefully controlling their nucleation and growth to produce a batch of particles very similar in size and shape. They then purify the solution, filtering out some 90 percent of the particles and keeping only those of the right dimensions. With this technique, Murray says, they can obtain collections of 10-nanometer particles that vary in size by no more than plus or minus 5 percent.
Murray and Sun have developed a process whereby a single layer of these particles self-assembles on a substrate, creating a thin film composed of grains almost identical in size and shape. By studying its properties, they can discover the strengths and weaknesses of the types of magnetic materials that may be put to work on hard disks in the future. Collaborative research at Almaden and Watson has demonstrated that stable ferromagnetic films can be grown with grains as small as 4 nanometers. Such results support the possibility of producing storage media with much smaller grains than are now feasible.
The ultimate magnetic recording disk, Rosen says, would be a "patterned" disk, one containing a uniform array of bits, each a single grain of magnetic material instead of hundreds of grains. The grains would have to be large enough to maintain their magnetization in the face of temperature changes. Owing to its embryonic state, says Rosen, "this kind of technology is not around the corner." But it would enable magnetic disks to store a terabit -- a trillion bits -- per square inch. That's almost 30 times the greatest density that has been demonstrated to date. And if magnetic storage continues improving at its current rate, such disks will arrive in less than a decade.
Toward carbon circuits
Eventually, no matter how many improvements researchers make in the design of the computer chip, the conventional silicon-based CMOS scheme will have to be replaced. It may be 10 years, it may be 20, but everyone in the field realizes that at some point the technology will reach fundamental limits. No one knows what approach will supplant it, but Phaedon Avouris at Watson believes that carbon nanotubes are a promising candidate.
"Nothing compares with the properties of nanotubes," Avouris says. They are very light, consisting mainly of open space -- yet their mechanical strength is 10 times that of steel. And they are excellent conductors, able to carry current densities as much as 100 times higher than copper. Depending on their structure, nanotubes can act as either metals or semiconductors. They also conduct heat very well, and, since carbon is chemically similar to silicon, it may be possible to build hybrid devices that are part silicon, part carbon nanotube.
Nanotubes are, as their name suggests, tiny tubes of pure carbon. In size, they range from 1 to 3 nanometers, or less than a tenth the size of the smallest features on today's silicon chips. The structure of a nanotube is much like a single sheet of graphite, with carbon atoms arranged in a hexagonal, honeycomblike structure, but the sheet is curved around to form a seamless tube. A nanotube's conductivity and other properties vary depending on how wide it is, whether the sheet was twisted when it formed the tube, and whether the tube is attached to a surface that distorts its shape (a condition that allows the nanotube to be modified externally). In theory, then, nanotubes could prove useful in a wide range of applications.
Already Avouris and his co-workers have created a field-effect transistor -- the basic component of computer chips -- with a single 1.5-nanometer-wide nanotube serving as the channel through which electrons flow from the source to the drain. By varying an external electric field, they were able to manipulate the current in the channel by five orders of magnitude (as is done with normal silicon transistors) even without trying to optimize the performance of the nanotube transistor.
But for Avouris, trying to build practical circuits would be premature. He is still exploring how electrons move through the nanotubes, what the intrinsic electrical resistance of the tubes is, and how external fields and conditions modify their properties. "At this point it's more important to understand the properties than to jump in and try to come up with devices," he says.
Theis believes this is the sort of approach that eventually pays off, usually in ways no one can predict.
"Experience has taught that the road to the future is paved with new materials -- for chips, displays,
magnetic disks and any other computer component," he says. If the information industry is to keep moving
forward as it has for the past several decades, it will only be by remembering the importance of the physical
stuff that embodies that nonmaterial information.
Robert Pool is a freelance science and technology writer based in Tallahassee, Florida. His most recent book
is Beyond Engineering: How Society Shapes Technology.
SIDEBAR: Computing a better LED
