Modeling magnetic materials
This section shows the final configuration of magnetic moments in five layers at the interface between iron-manganese (red and purple balls with yellow arrows showing magnetic moments) and cobalt (blue balls with yellow arrows.) The simulation of 2016 atoms had 15 iron-manganese and 6 cobalt layers.
For many years, information technology has been marked by an exponential growth in the
ability to store and retrieve data—so much so that the industry is reaching its limits on
further improvement in data storage and retrieval with the materials at hand. Developing
new materials requires research at the molecular and atomic level to understand their
properties. Such studies require calculations involving even the complexities of quantum
physics. Somewhat ironically, the ability to advance this type of research has, until
recently, required computational power that has been constrained by the very limits the
research is trying to overcome.
Yet the demand to handle more data faster keeps growing—for instance, stuffing more
data into smaller devices, such as digital cameras, and retrieving that information faster.
Meeting this demand requires creating new materials when the demands on current
materials are more than they can effectively meet. These are tough materials sciences
problems whose complex solutions will depend on calculations made using the latest and
most powerful supercomputers.
ORNL’s Malcolm Stocks, Thomas Schulthess, Xiaoguang Zhang, Don Nicholson, Bill
Shelton, and the University of Tennessee’s Balazs Ujfalussy, along with other
collaborators, are performing complex materials science calculations using ORNL’s
growing power in supercomputing. “We have been developing computational methods to
model magnetic states of various materials and to try to unravel the intricacies of
magnetism,” says Stocks.
Stocks, Shelton, Schulthess, and Ujfalussy have been exploring the role of
antiferromagnetic materials in contact with ferromagnetic materials in magnetic multilayer
storage and read-head devices, or, simply put, the tiny yet complex components that make
a computer’s hard drive tick. Data are stored, in binary fashion, through multitudes of
yes/no or on/off commands captured in the magnetic orientation of small regions of
magnetism, which are, in turn, made up of the magnetism of thousands of atoms. Individual
atomic “magnetic moments” result from an imbalance between the number of spin “up”
and spin “down” electrons associated with the atomic site. If the atomic magnetic
moments all point in the same direction, that material is ferromagnetic. (“Ferro” connotes
iron, the most common magnetic element.) It will stick to a refrigerator. It will make an
electric motor run. Or it can be used to store data.
In antiferromagnetic materials, the individual atomic magnetic moments are oriented in
the materials’ structure in such a way that the total magnetic moment is zero; for example,
the magnetic moments associated with individual atoms in a crystal can point “up” and
“down” in a regular manner, with as many “up” moments as “down” moments.
The read head in a computer disk drive is a multilayered component, known as a spin
valve (because it functions similarly to the way a mechanical valve might turn water on or
off). A spin-valve device consists of a ferromagnetic material, a nonmagnetic layer, and
another ferromagnetic material that is held constant by an adjoining layer of
antiferromagnetic material. Through an effect called “exchange bias” the magnetic
orientation of the ferromagnetic layer closest to the antiferromagnetic layer is held
constant, or “pinned,” enabling the other, outermost ferromagnetic layer to “switch” from
one orientation to the other, depending on the bit of information that is stored on the disk.
When this process cycles several million times, the word-processing document,
spreadsheet, or digital photo stored a few days ago appears on a computer screen.
To develop lighter, smaller devices for faster access to and storage of more data, new and
better materials are needed. Stocks, Schulthess, and colleagues at ORNL and across the
country have been exploring the incredibly complex interactions at the atomic levels that
can produce insights into the development of new, advanced materials for the next
generation of information technology.
Using computer simulations performed at ORNL, Stocks, Schulthess, Shelton, and
Ujfalussy have recently explored magnetism in iron manganese (FeMn), an alloy used as
the anitferromagnet in spin-valve devices. FeMn is an antiferromagnetic material used to
fix magnetic moments in those pinned layers of ferromagnetic material, to allow detection
of differences in electrical resistance (between the pinned and moving layers) that
represent 1’s and 0’s sensed on magnetic media. In addition, they are studying the
magnetic structure of interfaces between FeMn—an antiferromagnet—and cobalt (Co)—a
ferromagnet—to shed light on the mechanisms responsible for exchange bias.
In a collaboration with the Department of Energy’s National Energy Research Scientific
Computing Center (NERSC) at Lawrence Berkeley National Laboratory, the Pittsburgh
Supercomputer Center (PSC), and DOE’s Center for Computational Sciences (CCS) at
ORNL, the ORNL team has run simulations of FeMn and FeMn-Co interfaces at
unprecedented computational speeds. Calculations involving 2016 atoms ran at a
computational speed of 2.26 teraflops (2.26 trillion additions, subtractions,
multiplications, and divisions per second). The calculations involved 126
sixteen-processor nodes on the NERSC IBM SP3 supercomputer. A subsequent job
involving 2176 atoms ran at 2.46 teraflops on 136 nodes; this run was done at about 75%
efficiency, a figure much higher than has been seen in typical large-scale production
What have these computational experiments
shown? First, the FeMn alloy has a different
magnetic structure than previously thought.
Second, at interfaces between ferromagnetic Co
and antiferromagnetic FeMn, the magnetic
structure of the antiferromagnetic material
changes radically from that associated with bulk
Stocks explains that, previously, the magnetic
structure of FeMn was thought to be one of three
states of antiferromagnetic ordering, called 1Q,
2Q, and 3Q. This system was studied experimentally by various techniques, including
neutron scattering, but experiments were unable to unambiguously distinguish among these
three possible magnetic structures. The 3Q structure was thought to have the lowest
known energy and, therefore, to be the most stable, when compared with the 1Q and 2Q
structures. However, Bill Shelton found, through hours of high-performance computer
modeling using the 176-node IBM RS/6000 SP machine at ORNL, that a magnetic
configuration exists that is even more stable than the 3Q structure.
“We did calculations using constrained density functional theory and spin dynamics,”
Stocks says. “Our goal was to understand the alloy’s noncollinear, antiferromagnetic 3Q
magnetic structure. We predicted there is a relaxed, even lower-energy, and therefore
more stable, magnetic state that we call 3QR, in which R stands for relaxed. According to
our simulation, in this state, the magnetic moment orientations are pointed in slightly
different directions from those of the 3Q structure. We thought that this insight could be an
important ingredient in the explanation of exchange bias.”
Schulthess explains that with read heads, for instance, issues of practicality stand in the
way of improving their performance and further miniaturizing them. Improvements in the
spin-valve function could be obtained by creating a larger magnetic field to fix the
magnetic moments of the ferromagnetic layer. But exposing components to such a magnetic
field would likely also wreck the disk and the data stored on it. The other solution is to
increase the amount of antiferromagnetic material to enhance “the magic hand that
holds”—or pins—the magnetic moments of the ferromagnetic material. However, doing
so with currently available materials would require disproportionately more
antiferromagnetic material, thwarting any further attempts at miniaturization.
The ORNL team believes the key to making an improved, miniaturized spin-valve device
could lie in changes in the magnetic moments of the antiferromagnetic and ferromagnetic
interfaces, which surfaced in the computer models. “Details of the magnetic structure
appear to be very rich and could not have been predicted without the full machinery of
first-principles spin dynamics and the use of massively parallel computing resources,”
Stocks says. “Although the origins of exchange bias are currently not fully understood, it
is known that the magnetic structure of the ferromagnetic-antiferromagnetic interface is a
key ingredient. The ORNL group’s prediction of disorder in the magnetic moment
orientation and the large changes that occur when the antiferromagnet is proximate to a
ferromagnet have a number of interesting implications for the mechanism controlling
“It’s all in how the magnetic moments arrange themselves,” says Schulthess.
Stocks, Shelton, Schulthess, and Ujfalussy performed their calculations using a program
called the Locally Self-Consistent Multiple Scattering (LSMS) code, which was
developed at ORNL on the Laboratory’s first supercomputer, the Intel Paragon. The
45,000-line LSMS code continues to be developed and extended to treat larger and more
complex systems. Currently, the code is modeling magnetic materials on ORNL’s latest
supercomputers, including the Lab’s new IBM Power4 machine at CCS. The researchers
have run their models on the most powerful computers they can line up, including
supercomputers at NERSC and PSC. Supercomputing high points came when they won the
prestigious Gordon Bell Prize in 1998 and were the first to run a real application code at
a speed of greater than 1 teraflop. Most recently, the code ran at a staggering 4.46
teraflops on a Compaq supercomputer at PSC.
Advances in supercomputing and advances in materials technology could develop into a
symbiotic and self-perpetuating relationship. Strides in one area may very well lead to
strides in the other. Supercomputers are helping materials researchers understand what’s
happening with individual atoms, as well as the incredibly complex relationships that
exist between those atoms and at the interfaces of the materials they compose. As both the
researcher and the casual Web surfer can already attest, the sky may be the limit for what
research into new materials for computers can produce and the speed and amount of
information the resulting computers can process.—Bill Cabage, editor of the ORNL
Reporter, an employee newsletter.