Inverse Design of High-Dimensional Nanostructured 2x2 Optical Processors Based On Deep Convolutional Neural Networks

Optical computation, especially the optical neural network, has gained great attention recently due to their high computation throughput and low energy consumption. The 2x2 optical processor, as a building block of high-order matrix multiplication circuit, can be an essential part for optical neural networks. For the first time, we demonstrate a compact and general optical processor on silicon-on-isolator platform with a footprint of 1.68x3.6 mu m(2). The transfer matrix of an optical processor is determined by its nanostructured pattern, which has 2(420) possible combinations. To accelerate the design process of arbitrary optical processors, a two-step trained tandem model consisting of a forward model and an inverse model based on deep convolutional neural networks is proposed. After training, the forward model is able to predict the transfer matrix for a given optical processor with prediction accuracy of 98.8%, while the calculation speed is more than one thousand times faster than the electromagnetic simulation. The inverse model can predict the geometry of an optical processor for a target transfer matrix with prediction accuracy of 96.5% and its prediction time is also within a second. This two-step trained tandem model paves a new way for optical processors design.

Automated summary

Optical computation, particularly optical neural networks, have received attention due to their high throughput and low energy consumption. A compact and general optical processor was demonstrated on a silicon-on-isolator platform, and a two-step trained tandem model based on deep convolutional neural networks was proposed to accelerate the design process of optical processors.