By J. William Bell, NCSA Senior Science Writer -- They're researchers stuck in the middle -- between theories of varying precision, between models of varying size, even between the worlds of theoretical and experimental chemistry. Like a farm cut into a mountain foothill, it's a challenging but fertile row to hoe. The team, out of Stanford University and headed by Charles Musgrave, studies the chemical reactions that occur on the surface of microchips during their manufacture. By looking at the reactions at the atomic level, team members hope to illuminate the basic science that will improve production and perhaps lead to novel applications and devices. "Attempts to computationally prototype the reactors used to manufacture computer chips have really only been around for about 10 years," says Musgrave, an assistant professor of chemical engineering and materials science. "In order to create these prototypes, you have to be able to model and change all the elements of the reactor used to build the chip, but you also need to fully understand the chemistry at the chip's surface. That's where we come in." With the help of an Alliance SGI Origin2000 supercomputer at NCSA, two members of the team -- PhD candidates Collin Mui and Yuniarto Widjaja -- are becoming particularly "expert at predicting the energetics and behavior of the species that populate the surfaces," according to Musgrave. Chip Chemistry A computer chip begins its life as simple quartz (SiO2), that transparent rock underfoot every time you hit the beach. Quartz is easy to come by -- it makes up about a quarter of the planet's crust -- but tough to make useful to manufacturers. Only a handful of companies around the world process quartz into the pure silicon used as the chip's base. To conjure up the purification process, imagine the smelting heat of a foundry coupled with the precision of a surgical theater. Chip-quality silicon is heated to more than 2,500 F, and only defects less than 1/300th the width of a human hair are acceptable in the silicon ingots that are cut into wafers. As a semiconductor, silicon can be manipulated into either a conductor or an insulator by selectively introducing different impurities. These intricate patterns of conducting and insulating silicon form the basis for the complex circuits of the chip, which are built on the silicon wafer by depositing and removing different materials. More than 20 layers of deposited films are used to complete the three-dimensional structure of the chip's circuitry. Metals such as copper and aluminum are introduced as conductors to wire the chip, and various reactants are used to create insulators. The team starts their models -- just like a chip -- with a silicon surface, modeled as clusters of nine, 21, or 33 silicon atoms. Bonds that are broken when extracting this virtual chunk of the silicon surface are "terminated" with hydrogen atoms. Adding hydrogen atoms creates a slightly different chemical environment than the pure silicon of a chip. But not adding the hydrogen causes a modeling problem that's more difficult to address. Unterminated, hypothetical clusters of atoms would have different hybridizations than the pure, real-world silicon; attaching hydrogen in the right places allows researchers to avoid this issue. The goal is to simulate reactions on the small virtual surface of the cluster. If successful, the simulated reactions will mimic reactions on a wafer in a modern semiconductor fabrication plant. Becoming expert The Stanford team is currently focused on how and under what conditions reactants break down and how components incorporate themselves into the chip surface. Most of their models rely on density function theory (DFT) calculations, which solve the Schrödinger equation that governs the behavior of subatomic particles in molecules. Widjaja, for example, studies the behavior of ammonia during deposition. Ammonia (NH3) is known to be adsorbed -- retained on the cluster's surface without penetrating its bulk -- on silicon and to dissociate into NH2 and a hydrogen atom. Designers aren't much interested in that lonely hydrogen on the surface, however. In deposition of ammonia on silicon, they are really interested in the nitrogen that ammonia offers. Nitrogen reacts with the silicon to produce silicon nitride, which is commonly used as an insulator, oxidation mask, or diffusion barrier on chips. The unwanted hydrogen atom, on the other hand, pairs with a neighboring a hydrogen atom and desorbs. Chemical engineers know that if you raise the temperature in the reactor to between 700 and 1,000 K, the NH2 will further decompose, hydrogen will desorb from the silicon surface, and silicon nitride will form. The details of how this happens aren't known. The nature of the chemical species formed, the mechanisms that cause the formation, and the precise energies required to trigger those mechanisms have gone largely unexplored. Widjaja's models give some of the first atomistic views of those features and events. Mui, meanwhile, looks at methylamines, organic ammonia derivatives in which methyl groups (CH3) replace ammonia's hydrogen. Though it is more thermodynamically favorable for the methyl groups to break from the nitrogen before the hydrogen does, the models show that the hydrogen-nitrogen bonds break first. The models also show that trimethylamines, which have only methyl groups and no lone hydrogens, do not dissociate at all because there are no hydrogen-nitrogen bonds. By understanding the basics of this interaction -- and lack of interaction -- new applications that exploit previously ignored materials may be developed. "Organics like methylamines currently aren't used in industry, so we're looking at how they react. We want to build up a knowledge base. Find out what might be useful. What is even possible." Mui says. Ultimately, this knowledge might allow designers use more complex reactants to build more complex structures, such as molecular switches and sensors, and nanotechnology machinery. Quantum chemical techniques are essential to exploring these systems. Farther than fingerprints The nature of these highly specialized models makes it difficult, if not impossible, to confirm most of the findings through laboratory trial. "A big effort goes into the entire mechanism, showing all the fundamental steps involved, all the pathways," says Musgrave. "[Experimentalists] often get the first steps, but we get chemistry that's most often not attainable experimentally." Adds Mui, "At this level, experiments give you fingerprints of the reactions' products, but they don't tell you why you get those products. Why the reaction takes place. Why the reaction stops or continues." "Experimentation allows you to see. Theory explains what you saw and even lets you predicts what are you going to see," he says. "It is the ultimate microscope in that it lets you study atomic-scale systems in great detail, even if they haven't yet been made." The team bridges the gap between modeling and experimentation in at least two ways. First, there's constant interaction and collaboration with chemists who specialize in more traditional research methods. Mui, in fact, splits his time between Musgrave's team and the lab of another member of Stanford's faculty, Stacey Bent, an assistant professor of chemical engineering. As with Musgrave's team, her research team focuses on fundamental chemistry -- including the mechanisms, kinetics, and reaction pathways of product growth -- under the extreme heat and pressure that mark chip production. But Bent's research does keep its feet firmly planted in situ and uses a variety of spectroscopic techniques. Second, the team spends a portion of their NCSA computing time modeling with a method, known as quadratic configuration interaction (QCI), that is generally accepted as more accurate than DFT. This accuracy is often needed to predict properties like chemical kinetics that require more precision. Unfortunately, QCI is too computationally expensive for general use. "The smallest model takes about 20 to 30 hours with DFT on a single workstation. With QCI it could take 3,000 hours. The largest models we look at would take years with QCI," Widjaja says. Using QCI sparingly on the Origin2000 gives the team the freedom to run some of these otherwise intractable models. The results allow team members to substantiate the accuracy of the DFT models. "With just DFT, you're done, but you don't know if you're right, especially when you're working in an area that doesn't have an experimental base. With the Origin2000 runs, we've gained a lot of confidence in our results," Musgrave says. Perhaps it's not the same level of confidence that an experimental confirmation would afford. But for researchers who revel in being stuck in the middle, it's a great start. It's just a matter of waiting for the experimentalists to catch up. Relevant URLs: --Charles Musgrave's homepage http://chemeng.stanford.edu/html/musgrave.html ---------- Story re-printed courtesy of NCSA’s Access Online ----------