Optimizing Mach-Zehnder Interferometer Mesh for Photonic Computing

Tuesday - October 3 2023, 18:26 UTC - 1 year ago

Huazhong University of Science and Technology's team has developed a Mach-Zehnder interferometer mesh that is more efficient at performing real-valued matrix-vector multiplication than the conventional meshes. This mesh is more efficient because of the fewer phase shifters it requires and finds application in large-scale optical neural networks also.

As research into photonic computing progresses, scientists seek to optimize the performance of optical computing devices by making purpose-specific changes to their design. A team led by Bo Wu and Shaojie Liu at Huazhong University of Science and Technology in China designed and tested a kind of Mach-Zehnder interferometer mesh that is more efficient at performing real-valued matrix-vector multiplication. Their research was published Sep. 19 in Intelligent Computing.

Conventional Mach-Zehnder interferometer mesh has many advantages, including large bandwidth and high stability. However, it was designed for complex-valued matrix-vector multiplication, a type of computation that has hardware requirements different from those of real-valued matrix-vector multiplication. The new mesh is customized to take advantage of differences between the two types of computation.

The new mesh is more efficient because real-valued matrices have half as many degrees of freedom as complex-valued matrices, and thus require half as many phase shifters. Fewer phase shifters means a smaller piece of hardware that requires less electricity to operate. Moreover, the simplified mesh is easier to use and easier to fabricate because of the way it detects light.

The conventional mesh requires "coherent" light, such as light from a laser, whose waves are all the same size, shape, and direction. The simplified mesh detects "incoherent" light, a kind of light whose waves are much less regular and can be produced by a wider variety of light sources.

According to the researchers, "The results show that the proposed MZI mesh exhibits excellent performance in the benchmark task. Additionally, the error analysis indicates that the proposed scheme is robust to fabrication errors." Moreover, the simplified mesh is scalable, which means it holds great potential for use in large-scale optical neural networks, where the savings wouldus/robotics/mach-zehnder-interferometer-mesh-for-photonic-computingbe similarly large.

"We first designed the architecture with our intuition and common sense because there is a clear redundancy of phase shifters," said corresponding author Hailong Zhou of Huazhong University of Science and Technology.

The mesh was developed in several steps, beginning with a "type 0" conventional mesh for a complex-valued optical matrix, progressing to a "type 1" mesh with fewer phase shifters, then to a "type 2" mesh with an additional output port, then to a "type 3" mesh with a smaller footprint. The effectiveness of the type 3 mesh is supported by mathematical proofs and numerical simulation. "We just wanted to take a chance by doing the numerical simulation but surprisingly got an excellent result." .

The simulation was the popular iris classification benchmark task, in which a computer must "learn" the properties of three different types of iris flower by processing a dataset of physical measurements corresponding to examples of each type. The training algorithm used in this research was particle swarm optimization. The next step, the authors said, is to perform a large-scale image classification task.

Optical neural networks are hardware devices that serve the same function as artificial neurally-controlled machines. Just as traditional neural networks operate via layers of nodes, each of which performs some processing of data, optical neural networks use optical waveguides to achieve a similar effect.