Deep Learning Enabled Design of Complex Transmission Matrices for Universal Optical Components

Abstract

Recent breakthroughs in photonics-based quantum, neuromorphic and analogue processing have pointed out the need for new schemes for fully programmable nanophotonic devices. Universal optical elements based on interferometer meshes are large compared to the limited chip real estate, restricting the scalability of the approach. Here, we propose an ultracompact platform for low-loss programmable elements using the complex transmission matrix of a multi-port multimode waveguide. Our approach allows the design of arbitrary transmission matrices using patterns of weakly scattering perturbations, which is successfully achieved by means of a deep learning inverse network. The demonstrated platform allows full control over both the intensity and phase of all outputs in a 3x3 multiport device using a footprint of 33x6 µm 2 and for typical perturbations achievable in experiments.

Nonvolatile programmable silicon photonics using an ultralow-loss Sb 2 Se 3 phase change material

“…MMIs are of interest for their small footprint as well as for their substantially reduced sensitivity to the environment when compared to MZIs. The device principle is based on the concept of wavefront shaping, where a distribution of weak perturbation pixels is used to effectively steer the light with very low loss ( 22 , 23 ). The perturbation induced by switching of the PCM layer is sufficient to apply this device concept in a practical scalable device configuration.…”

“…The results presented here unequivocally show the capabilities of the new phase change technology in providing low-loss programmable optical phase control in PICs. This approach is expected to open many new applications in postfabrication device tuning, programmable weight banks, unitary matrix operations, and, ultimately, all-optical field-programmable arrays ( 10 , 23 ). Nonvolatile PCM-based approaches may hold an energy advantage compared to active devices, which require driving voltages to maintain a configuration and hence could offer opportunities for reduced power consumption PICs.…”

“…Figure 1B shows a schematic layout of this experiment and highlights our gen eral approach of optical writing of pixels using a microscope. Fol lowing on earlier proofofprinciple studies (22,23), the capability of writing nonvolatile pixel patterns in a multiport waveguide signi fies a breakthrough in freeform control over the flow of light inside integrated photonic circuits. Currently dense meshes of singlemode devices are being considered as the main strategy for programmable optical circuits; however, this is expensive in both the number of components and device footprint.…”

“…Currently dense meshes of single-mode devices are being considered as the main strategy for programmable optical circuits; however, this is expensive in both the number of components and device footprint. The results presented here are a first step toward a new family of programmable multiport elements, which could open up entirely new architectures for photonic integrated circuits ( 23 ).…”

Interconnect-Free Multibit Arithmetic and Logic Unit in a Single Reconfigurable 3 μm2 Plasmonic Cavity

“…S7 for details), the implementation of n-bit ALU combining Boolean and arithmetic functions will require a global parametric optimization for which photonic inverse design and evolutionary tools could be applied in order to maximize the output intensities and locations and the signal-to-noise ratio. 27,37,38,40 Yet, akin the analogous computing based on metasurfaces, 28 the main challenge towards increased logic complexity resides in the inverse design of the adequate plasmonic cavity to reach a specific ALU function for which more advanced deep learning [41][42][43] approaches are foreseen to provide powerful tools.…”