Neural Network Design of Multilayer Metamaterial for Temporal Differentiation

Abstract

However, physical limitations (such as quantum tunneling) imposed on the size and separation of individual transistors in semiconductor-based circuits means that performance gains may soon reach their maximum potential. [2] Additionally, parasitic capacitances appearing in these devices can lead to challenges in energy consumption and speed during the charging/discharging processes involved when using them as switching devices for computing. [3] While it is expected that semiconductor-based technology remains with us for many years to come and research is underway to overcome challenges and further improve its performance, ever-increasing demand for faster and more energy efficient computations therefore necessitates the introduction of alternative computing paradigms. Some notable examples that have been explored so far include quantum computing, [4] TEM pulse switching, [5][6][7][8] and computing with solitons, [9] to name a few. In this realm, analogue wave-based computing, as an alternative to current electronic-based computing systems, offers inherently parallel processing capabilities for fast, energy efficient computation on specialized tasks such as high throughput image processing, [10,11] equation solving, [12][13][14] and machine learning. [15] Another prominent area of research where this paradigm shift is taking off is in the field of metamaterials (MTMs)and metasurfaces (MTSs) as their 2D counterparts. [16,17] MTMs and MTSs are artificially engineered structures that enable enhanced and tailored wave-matter interactions both in space and time [18][19][20][21][22] with responses not easily available in nature such as epsilon near-zero or negative refractive index. [17,[23][24][25] For the past 2 decades, the scientific community has reported many interesting applications using MTMs across different branches of science and engineering. For example, MTM-based antennas for high gain 5G communications, [26] miniaturized optical neural networks (NNs) for real-time object recognition, [27] MTS-based sensors at terahertz frequencies [28] and invisibility cloaking devices, [29] to name a few. Recently, the concept of wave-based analogue computing with MTMs has also been introduced. [30] In this emerging branch of MTM research, the interactions between an incoming wave and the multiple metaatoms (scatterers) can be exploited to perform some mathematical operation of choice (e.g., differentiation, integration, Controlling wave−matter interactions with metamaterials (MTMs) for the calculation of mathematical operations has become an important paradigm for analogue computing given their ability to dramatically increase computational processing speeds. Here, motivated by the importance of performing mathematical operations on temporal signals, multilayer MTMs with the ability to calculate the derivative of temporally modulated signals are proposed, designed, and studied. To do this, a neural network (NN) based algorithm is used to design the multilayer structures (alternating layers of indium tin oxide (ITO) and ...

Machine Intelligence in Metamaterials Design

“…ANNs solve complex problems, such as identification, classification, or recognition with high accuracy and at high speed. The ANN analyzes input data to make accurate decisions [51].…”

Complex Matrix Equation Solver Based on Computational Metasurface

“…The first type is based on Fourier transform methods. [4][5][6][7][8][9][10][11][12][13] The general architecture of this type consists of artificial lenses like metamaterials or metasurfaces configured in 4F form and provides certain mathematical relationships between the input and output ports. However, most of the inputand-output relationships can only be differential or integral due to the limitation of fundamental physics, which greatly restricts such systems from obtaining a wider range of applications.…”

Time derivatives via interconnected waveguides

“…In this realm, different examples of EM wave-based computing structures have been recently reported such as optical networks able to perform computing operations such as matrix inversion 12 – 15 , transverse electromagnetic (TEM) pulse switching with waveguide networks 16 – 19 and analogue computing with dielectric multilayers 11 , 20 . Furthermore, the introduction of metamaterials 21 , 22 , artificial media which can exhibit exceptional control over waves in space and time 23 – 31 , has led to the concept of “computational metamaterials” first introduced in 2014 by Silva et al 11 .…”

“…Different EM wave-based analogue processors have been reported performing first order differentiation, in both space and time domains, with examples including structures designed by tailoring the permittivity distribution or reflection/transmission spectra of a metamaterial block/metasurface 9 , 32 – 34 , 38 , 39 . In practice, this often requires the fine tuning of several design parameters, such as the lengths of dielectric layers in a multilayer structure or the permittivity of a pixel in a 2D grid 9 , 11 , 20 . To achieve this goal, various design techniques have been recently applied and demonstrated such as fiber gratings 40 , 41 , Mach–Zehnder interferometers 42 , advanced optimization and inverse design 43 , 44 and also machine learning approaches 20 , 45 , 46 .…”

“…In practice, this often requires the fine tuning of several design parameters, such as the lengths of dielectric layers in a multilayer structure or the permittivity of a pixel in a 2D grid 9 , 11 , 20 . To achieve this goal, various design techniques have been recently applied and demonstrated such as fiber gratings 40 , 41 , Mach–Zehnder interferometers 42 , advanced optimization and inverse design 43 , 44 and also machine learning approaches 20 , 45 , 46 .…”