Nathaniel Kinsey, Ph.D., Engineering Foundation Professor in the Department of Electrical and Computer Engineering (ECE), is leading a group to bring new relevance to a decades-old computing concept called a perceptron. Emulating biological neuron functions of the messenger cells within the body’s central nervous system, perceptrons are an algorithmic model for classifying binary input.
When combined within a neural network, perceptrons become a powerful component for machine learning. However, instead of using traditional digital processing, Kinsey seeks to create this system using light with funding from the Air Force Office of Scientific Research. This “nonlinear optical perceptron” is an ambitious undertaking that blends advanced optics, machine learning and nanotechnology.
“If you put a black sheet outside on a sunny day, it heats up, causing properties such as its refractive index to change,” Kinsey said. “That’s because the object is absorbing various wavelengths of light. Now, if you design a material that is orders of magnitude more complex than a sheet of black plastic, we can use this change in refractive index to modify the reflection or transmission of individual colors - controlling the flow of light with light.”
Refractive index is an expression of a material’s ability to bend light. Researchers can harness those refractive qualities to create a switch similar to the binary 1-0 base of digital silicon chip computing. Kinsey and collaborators from the U.S. National Institute of Standards and Technology, including his former VCU Ph.D. student Dhruv Fomra, are currently working to design a new kind of optically sensitive material. Their goal is to engineer and produce a device combining a unique nonlinear material, called epsilon-near-zero, and a nanostructured surface to offer improved control over transmission and reflection of light.
Kinsey’s prior research has demonstrated that epsilon-near-zero materials combine unique features that allow their refractive index to be modified quite radically - from 0.3 to 1.3 under optical illumination - which is roughly equivalent to the difference between a reflective metal and transparent water. While an effective binary switch, the large change in index requires a lot of energy (~1 milli-Joules per square centimeter). By combining epsilon-near-zero with a specifically designed nanostructure exhibiting surface lattice resonance, Kinsey hopes to achieve a reduction in the required energy to activate the response. The unique response of a nanostructure exhibiting surface lattice resonance allows light to effectively be bent 90 degrees, arriving perpendicular to the surface while being split into two waves that travel along the surface. When a large area of the nanostructure is illuminated, the waves traveling along the surface mix, where they interfere constructively or destructively with each other. This interference can produce strong modification to reflection and transmission that is very sensitive to the geometry of the nanostructure, the wavelength of the incident light and the refractive index of the surrounding materials. The mixing of optical signals along the surface can also selectively switch regions of the epsilon-near-zero material thereby performing processing operations.
A key aspect of Kinsey’s work is to build nonlinear components, like diodes and transistors, that use optical signals instead of electrical ones. Transistors and other traditional electronic components are nonlinear by default because electrical charges strongly interact with each other (for example, two electrons will tend to repel each other). Creating optical nonlinear components is challenging because photons do not strongly interact, they just pass through each other. To correct for this, Kinsey employs materials whose properties change in response to incident light, but the interaction is weak and thus requires large energies to utilize. Kinsey’s device aims to reduce that energy requirement while simultaneously shaping light to perform useful operations through the use of the nanostructured surface and lightwave interference.
The United States Department of Defense sees optical computing as the next step in military imaging. Kinsey’s work, while challenging, has potential to yield an enormous payoff.
“Let’s say you want to find a tank within an image,” Kinsey said, “Using a camera to capture the scene, translate that image into an electrical signal and run it through a traditional, silicon-circuit-based computer processor takes a lot of processing power. Especially when you try to detect, transfer, and process higher pixel resolutions. With the nonlinear optical perceptron, we’re trying to discover if we can perform the same kinds of operations purely in the optical domain without having to translate anything into electrical signals.”
Linear optical systems, like metasurfaces and photonic integrated circuits, can already process information using only a fraction of the power of traditional tools. Building nonlinear optical systems would expand the functionality of these existing linear systems, making them ideal for remote sensing platforms on drones and satellites. Initially, the resolution would not be as sharp as traditional cameras, but optical processing built into the device would translate an image into a notification of tanks, troops on the move, for example. Kinsey suggests optical-computing surveillance would make an ideal early warning system to supplement traditional technology.
“Elimination or minimization of electronics has been a kind of engineering holy grail for a number of years,” Kinsey said, “For situations where information naturally exists in the form of light, why not have an optical-in and optical-out system without electronics in the middle?”
Linear optical computing uses minimal power, but is not capable of complex image processing. Kinsey’s research seeks to answer if the additional power requirement of nonlinear optical computing is worthwhile given its ability to handle more complex processing tasks.
Nonlinear optical computing could be applied to a number of non-military applications. In driverless cars, optical computing could make better light detection and ranging equipment (better known as LIDAR). Dark field microscopy already uses related optical processing techniques for ‘edge detection’ that allows researchers to directly view details without the electronic processing of an image. Telecommunications could also benefit from optical processing, using optical neural networks to read address labels and send data packets without having to do an optical to electrical conversion.
The concept of optical computing is not new, but interest (and funding) in theory and development waned in the 1980s and 1990s when silicon chip processing proved to be more cost effective. Recent years have seen many advancements in computing, but the more recent slowdown in scaling of silicon-based technologies have opened the door to new data processing technologies.
“Optical computing could be the next big thing in computing technology,” Kinsey said. “But there are plenty of other contenders — such as quantum computing — for the next new presence in the computational ecosystem. Whatever comes up, I think that photonics and optics are going to be more and more prevalent in these new ways of computation, even if it doesn't look like a processor that does optical computing.”
Kinsey and other researchers working in the field are in the early stages of scientific exploration into these optical computing devices. Consumer applications are still decades away, but with silicon-based systems reaching the limit of their potential, the future for this light-based technology is bright.