In recent years, considerable progress has been made in the development of advanced sensors, capable of detecting increasingly complex signal patterns. Hyperspectral imagers, for instance, that acquire data resolved not only in space but also spectrally (i.e., at many wavelengths), enable the extraction of unique characteristics, unobtainable by other means. For example, looking at a ball containing blueberries and grapes, one can, with the same sensor, not only distinguish each individual fruit, but also estimate the freshness of each or whether it has humidity on its surface. The ongoing trend toward sensor miniaturization (down to nanoscale sensors) provides a strong incentive to develop computational capabilities for processing the rich signal information at the sensor's scale. A major step toward developing computational "brain" power to speed up the processing of signal patterns is being taken by a multidisciplinary nanotechnology project at ORNL, supported partly by internal funding from the Laboratory Directed Research and Development Program. Researchers at the Laboratory are fabricating a nanoscale pattern-recognition device, using gold nano-particles on a DNA template, which may eventually prove the feasibility of this concept.
Specifically, the ORNL team is designing a quantum-dot array that can be operated at room temperature to carry out innovative computations. The construction and operation of this array, with the help of special algorithms, will constitute the world's first successful use of a nanoscale device to solve nontrivial computational problems such as signal discrimination. The advantage of embedding such a nanoscale computer into a micro- or nanosensor is that it avoids the need to send the signal from the sensor to a conventional computer a long "distance" away (both literally and figuratively in terms of scale). Information can be processed at the site of the sensor. An array of specially coated and closely placed gold nanoparticles, called quantum dots, may enable the operation of such "smart sensors" at room temperature.
ORNL's first quantum-dot array, which will straddle gold electrodes, will receive a vector of input currents from a sensor and produce a vector of output currents. It will be "trained" to classify patterns into categories that characterize specific classes of properties to be detected by the sensor. The output currents will indicate the class of the observed pattern, to allow appropriate actions to be taken.
"Our goal is to demonstrate via a proof-of-principle experiment how to produce a nanoscale information-processing device for a sensor," says Jacob Barhen, director of ORNL's Center for Engineering Science Advanced Research (CESAR) and head of this project. Barhen, Yehuda Braiman, Vladimir Protopopescu, and Nageswara Rao, all from CESAR and ORNL's Computer Science and Mathematics Division (CSMD), are developing the methodology and algorithms needed to implement neuromorphic computations, which will allow the quantum-dot array to learn and retrieve information.
The practical goal of the ORNL team is to build a device that emulates a neural network. Instead of the neurons and connecting synapses found in the brain, the nanoscale computer will depend on electrically charged gold quantum dots connected by electrons that tunnel between them at different rates.
In a neural network, the learning process modifies the interconnection strengths (synapses) between neurons. In a quantum-dot circuit, the inverse of a generalized capacitance matrix plays the role of the synaptic matrix. Unfortunately, once fabricated, the elements of this capacitance matrix are essentially fixed. Braiman suggested that exposing the array to an excitatory electromagnetic field would result in a modification of the tunneling rates of single electrons hopping between the quantum dots. By modifying the tunneling rate, the device will produce different output current patterns for the given input signals into the array. Using this phenomenon of photon-assisted tunneling, Barhen designed a neuromorphic learning algorithm, in which the amplitudes and frequencies of the polychromatic excitatory field are used as control parameters to achieve maximal discrimination between various classes of signals. Braiman and Protopopescu are focusing on issues related to the stability of the nonlinear dynamics underlying the device's operation, while Rao is exploring more advanced architectures for digital signal processing using nanodevices.
"The microwave field provides degrees of freedom and the ability to change electron pathways and rates of electrical conduction," Barhen says. "In this way, this device will mimic neurons in the brain. The pathways that the electrons take to minimize the discrepancy between a desired pattern class signature and one produced by the array under excitatory field illumination will enable the solution of the pattern classification problem. The array will consist of gold nanoparticles 1.5 nanometers (nm) in diameter that self-assemble by attaching to pre-selected locations about 3.5 nm apart on a specially engineered DNA template about 70 nm long. The nanoscale size of the particles and their regular placement in close proximity to one another is necessary for the array to function as a nanoscale computer at room temperature.
Because of gold's affinity to sulfur, the gold particles are coated and stabilized, or "passivated," by sulfur-containing molecules attached to gold molecules that connect into the DNA template. Specifically, the coating is a monolayer of alkylthiol organic molecules terminated with a carboxyl group (containing carbon, oxygen, and hydrogen) that binds to amino groups, attached to the DNA at specified locations. (Amino groups, which contain nitrogen and hydrogen, used in the coupling to gold, are external to the DNA, in contrast to amines involved in the pairing of bases between two strands of DNA to make it double stranded.)
The material's fabrication and assembly are being performed by CESAR's Leon Maya, a chemist in the Chemical and Analytical Sciences Division (he makes the gold nanoparticles and coats them with the alkylthiol and carboxyl molecules), as well as CESAR researchers Karen Stevenson, Muralidharan Govindarajan, and Thomas Thundat, all of ORNL's Life Sciences Division, who attach the coated gold particles to the DNA template. So far, some 20 gold nanoparticles have been attached along a DNA template. The plan is to attach the gold nanoparticles in a grid on a DNA scaffold, which will be fitted between the electrodes. If necessary, the DNA bases between the gold particles will be destroyed using ozone or ultraviolet light, leaving the gold particles on a substrate spanning the space between the electrodes.
A gold nanoparticle is a cluster of gold atoms bound together by mutual attraction. In a cluster, most electrons circulate around the gold nuclei, but some hop back and forth between the outer shells of the gold atoms. Using the IBM supercomputer at DOE's Center for Computational Sciences at ORNL, CESAR's Jack Wells (of CSMD), working with David Dean and Mike Strayer (both of the Physics Division), is simulating the electronic density of clusters of gold atoms in support of the design of the quantum-dot array.
Understanding electron density in these clusters is important because the gold nano-particles arranged every 10 or 11 DNA bases on a DNA scaffold will conduct electricity in the same way that water drips from a faucet rather than as a steady water flow. If the voltage is high enough, the electrons flow by single-electron tunneling, hopping between the weakly coupled nanoparticles and producing a very nonlinear relationship between the current and the voltage. A range of low voltages could produce no current, representing a "0," and higher voltages could produce a current spike that represents a "1" for use in information processing.
Employing a functional density code obtained from the IBM Research Division in Zurich, Wells, in collaboration with Wanda Andreoni of IBM, has modeled a bare cluster of 38 gold atoms, calculated their electronic structure (the locations of the electrons in shells and between outer shells of the atoms), and related the size of a cluster to the number of its electron charges. He has also simulated the energy effects of adding or removing an electron.
"This information is sufficient to predict the electrical capacitance of the nanocluster and to allow us to quantitatively analyze structures in nanoscale circuits," Wells says. "Such information is difficult to obtain from direct experimental observation."
Wells modeled a passivated 38-atom gold cluster bound with 24 methylthiol groups (sulfur group plus a methyl group, or SCH3). "I found the binding with the chemical group causes the reorganization or rearrangement of the gold atoms in the cluster, compared to the idealized case of the unpassivated cluster," he says. These calculations being done to find new properties in very small features require a large amount of computing capacityabout one-third of the IBM supercomputer's nodes.
The early research success in attaching gold nanoparticles to a DNA template has been submitted for publication. The researchers are now taking on bigger challenges as they attempt to build a world-class nanoscale device to solve global problems.
Related Web sites
for Engineering Science Advanced Research (CESAR)