Abstract
Analog photonic solutions offer unique opportunities to address complex computational tasks with unprecedented performance in terms of energy dissipation and speeds, overcoming current limitations of modern computing architectures based on electron flows and digital approaches. The lack of modularization and lumped element reconfigurability in photonics has prevented the transition to an alloptical analog computing platform. Here, we explore, using numerical simulation, a nanophotonic platform based on epsilonnearzero materials capable of solving in the analog domain partial differential equations (PDE). Wavelength stretching in zeroindex media enables highly nonlocal interactions within the board based on the conduction of electric displacement, which can be monitored to extract the solution of a broad class of PDE problems. By exploiting the experimentally achieved control of deposition technique through process parameters, used in our simulations, we demonstrate the possibility of implementing the proposed nanooptic processor using CMOScompatible indiumtinoxide, whose optical properties can be tuned by carrier injection to obtain programmability at high speeds and low energy requirements. Our nanooptical analog processor can be integrated at chipscale, processing arbitrary inputs at the speed of light.
Introduction
Current computing tasks are characterized by an elevated level of complexity and therefore require high computational cost. A major criticality of today’s digital computing is related to the required computational power, which does not scale well with the problem complexity. For this reason, developing innovative analog accelerators, which could take the load off from traditional computers by solving specific complex processes, holds the promise to significantly reduce energy consumption and can lead to the development of nextgeneration heterogeneous computing systems. Analog computers are not a recent invention and are, in fact, wellrooted in human history prior to this digital age, being applied to a vast variety of fields^{1}. While digital processing became dominant in the preceding 50 years, recently, several research groups explore innovative solutions to address growing computelimitation to businessasusual. This trend favored the advancement of several nonVon Neumann hardware architectures, which aim to homomorphically map specific algorithms directly into hardware. Analog memory^{2,3}, neuromorphic photonics for deep learning applications^{4,5,6,7}, optical coprocessors for highspeed convolutions^{8}, integral equation solvers^{9}, and quantum analog computers^{10} are only some examples of currently explored analog architectures, which can tackle complex tasks more efficiently than a conventional digital processor.
One of the mathematical tasks that can exploit these paradigms and may greatly benefit from using analog coprocessors, and the object of this study, is solving partial differential equations (PDEs). In fact, numerous scientific and engineering problems require the solution of PDEs^{11}, such as problems in thermodynamics^{12}, aircraft design^{13}, and others in electrical and mechanical engineering field^{14}. For solving multidimensional PDEs, current processors require a large number of (iterative) operations, which are computationally intensive, and, based on the complexity, necessitate considerable amount of memory and power^{15}. Analog computers capable of solving PDEs have been implemented since the 1950s^{11}; primarily using networks or grids of resistive or reactive elements to model the spatial distribution of physical quantities such as voltage, current, and power (in electric distribution networks), electrical potential in space, stress in solid materials, temperature (in heat diffusion problems), pressure, fluid flow rate, and wave amplitude. However, the complexities of an effective integration of a highspeedprogrammable and concurrently energyefficient staticlike analog mesh significantly reduced the advancement of this technology.
Novel compute systems are often enabled by new hardware platforms, such as metatronics, which was introduced by Engheta a decade ago as the branch of optics that focuses on the control of light at the nanoscale through metamaterialinspired optical nanocircuitry. Thus, metatronics offers an avenue for designing dense circuits and analog compute paradigms. Despite finding diverse applications, such as ultrathin subwavelength filters for optical signal^{16,17}, circuit for performing mathematical operations^{9,18}, and for modeling Hamiltonian of quantum system^{19,20}, its application as an alternative integrated platform for implementing a compact analog reprogrammable processor is still outstanding. The applicability of the metatronic concepts has, so far, been limited by four main factors: (a) the lack of availability of materials required to realize its individual circuit elements, (b) the absence of controlled processes for straightforward fabrication, (c) lack of reprogrammability, “write”, and (d) difficulties in accessing the results, “read”.
Here, we introduce a nanoscopic programmable analog processor based on a metatronic nanocircuit board aiming to map a finite difference PDE onto a mesh similar to a network of resistors, but in an integrable and ultracompact fashion using nanooptics. We design and show through numerical simulation, an ultracompact nanooptic circuit design based on air groove meshes, engraved into indium tin oxide (ITO) substrate, operating at its epsilonnearzero (ENZ) wavelength. We show that such a metatronic circuit, mimicking an electrical lumped circuit, demonstrates the ability to solve PDEs using finite difference. We demonstrate, assisted by experimentally derived optical characterization and control of deposition techniques, that ITO nanoelements exhibiting resistive, inductive, and capacitive behavior, sputtered using different sputtering conditions^{21}, can be embedded within the air grooves. Electrostatically finetuning their optical properties at high speed^{22} through carrier accumulation/depletion enables approximate solutions to a variety of PDEs, including Laplace equation, Poisson equation, diffusion equation, and wave equation. Interestingly, the power consumption to obtain an approximate discretized solution is effortlessly computed within the transit time of the electromagnetic wave propagates in the chip, once the chip is set to a particular boundary condition. In this view, we ultimately envision the implementation of an alloptical readout paradigm based on a nearfield optical microscopy measurement, providing information of the local dielectric displacement, at given points of the nanooptics circuit, thus allowing to extract the results of the computation in a discretized manner. Such wavelengthscale analog compute nodes can offer a variety of functionality in compact and efficient signal processing, computing, or artificial intelligence.
Results and discussion
Metatronic circuit as a PDE finite difference solver
Before beginning the description of the nanooptic accelerator, we ought to describe a metatronic mesh and how it compares to a finite difference and lumped resistive meshes. (Fig. 1a) We start by considering the steadystate homogeneous Laplace’s differential equation, where f(x, y) is the distribution of the physical entity (i.e., PDE solution) of the twodimensional (2D) domain (i.e., mesh). As a relevant example, we select the homogeneous heat equation, because it is widely employed in a variety of mathematical and engineering problems, ranging from machine learning in graph Laplacian methods^{23} to image analysis^{24,25}. Then,
where the field f in the case of the heat equation denotes the temperature distribution, and can be written in the Cartesian 2D space via
The same equation may eventually represent a steadystate distribution of temperature, as well as stress distributions, potential, and flows. (Fig. 1b.i) The finite difference method allows current processors to iteratively solve differential equations, such as Eq. 2, by approximating them via a difference equation, in which finite differences approximate the derivatives. Considering equidistant points in the domain (and constant point function ∈_{i}), Eq. 2 can be approximated via the finite difference relaxation method (Fig. 1b), at the point O of a mesh, as^{11}:
The finite difference equation can be straightforwardly compared to the application of Kirchhoff’s law to the currents f(Pi) netting at the junction O of a lumped circuit mesh (Fig. 1b), where resistances are opportunely scaled (R = h^{2}). Ohm’s law, in fact, shows a linear relationship between voltage and current, whilst Kirchhoff’s law states that the input and output of current into a node will always be equal.
According to these equivalences, as previously proposed^{26,27}, approximate experimental solutions of Laplace’s and Poisson’s boundary value problems can be obtained using networks of electrical resistances with a rather low tolerance^{26}. The measurement of voltage values at grid points provides the equivalent discrete solution for given resistor values. However, when the parameters of the circuits are programmed at a given rate (i.e., write speed) and its inputs initiated, then an electric mesh might be affected by resistivecapacitive (RC) time delay. Moreover, at high speeds the time at which the signal changes become comparable to the propagation time in the circuit, which therefore makes it into a distributed network. Under these conditions, the circuit would lose the ability to map a finite difference meshgrid that is able to solve PDEs. For this reason, a processor capable of solving PDEs using nanooptics circuitry is highly desirable. Here, we exploit the fact that subwavelength nanoparticles (NPs) in the optical domain can be treated as lumped circuit elements, whose impedance is defined in terms of the induced displacement current j_{D}, in response to the local electric field E. According to Maxwell’s equations, \(\widetilde \varepsilon\), the complex material permittivity, relates these two quantities through
which, for element size considerably smaller than the optical wavelength (d << λ)^{28}, represents an equivalent Ohm’s law in the optical domain, enabling the mapping of the resistive circuit. Nevertheless, arranging different NPs with absolute control on permittivity and position is far from easy^{29,30}, and as the size of the circuit grows it is challenging to make sure that the flux of displacement current is funneled to ensure the required circuit connections without phase delays. For these reasons, ENZ substrates are ideal for the implementation of nanocircuit boards, as they enable light to travel through the grooves just like electric currents in copper wiring^{31} (Fig. 1b.iii). Resistors, capacitors, and inductors can be implemented within the air grooves by tailoring the local permittivity values. To map Eq. 3 in the metatronics paradigm, the resistors are modeled as dissipative dielectrics. Due to the confinement of displacement current in the air grooves^{32}, as in an ideal ENZ substrate there is no displacement current leakage in the background, the circuit elements are locally coupled, implying that the Norton/Thevenin equivalents are admissible. In other words, local variations in the network generate global effects. Therefore, for a limited functional bandwidth over which the material of the board is ENZ (proportional to the fullwidth at halfmaximum of the ENZ resonance peak), Kirchhoff’s law is satisfied in the mesh, resulting in identical results with respect to a resistive network, reported in Eq. 3.
A node of a metatronic mesh, equivalent to a resistive one, is represented in Fig. 1b. This approach is not limited to resistive networks and consequently to Laplace homogeneous equations. Indeed, by using plasmonic NPs (negative permittivity) to construct nanoinductors and dielectric nanoelements (positive permittivity) for nanocapacitors, it is possible to map other timedependent PDEs applied to a 2D mesh, such as diffusion or wave equation (Fig. 1c).
Programmability of the nanooptic elements
In order to prove the ability to reprogram the circuit and control its electromagnetic response^{33}, in this section, we aim to demonstrate that different lumpedelement functionalities can be achieved in ITO films by modulating the free carriers of the ITO nanoelements via electrostatic doping.
In detail, an electrical voltage bias can be applied across a capacitor, whose plate is the ITO film, and according to the bias polarity an accumulation/depletion layer is formed at the ITO–oxide interface, increasing/decreasing locally the ITO’s carrier density (1 × 10^{19} −1 × 10^{21} cm^{−3})^{34,35,36}, which ultimately alters its complex optical properties.
For doing so, we use a straightforward, yet effective, demonstration of a programmable nanooptic circuit based on ITO in which a metaring is engraved in an ITO substrate, defining an air loop (Fig. 2a). The optical response of the system is obtained by performing a fullwave electromagnetic simulation of an ITO layer. An electric field is driven by a dipole (quantum dot, QD) placed within the channel, forming a directional flow of the displacement current when the ITO layer is in the ENZ range. By way of an example, in the simulation, the QD is considered as point source, neglecting farfield to nearfield coupling. In this case, we optimized the air trench of width 0.1 λ to form a closed loop with an average total length of 2πr = 2440 nm which is twice the wavelength, and it is sufficiently small for supporting TE_{10}like mode excited by a dipole and concurrently does not experience phase variation within the trenches. In Supplementary Fig. 2a, b we provide the eigenmode analysis as function of the width and height of the air trenches. For this exemplaryselected condition, the characteristic length scale of the nanooptic chip (D) is considerably smaller than the effective wavelength, therefore guaranteeing quasistatic approximation (D < λ_{eff}, D < λ, πD ~ λ_{eff}).
It is worth noticing that suitable QDs for these wavelengths are scarce^{37} and usually comparable in size with the trenches. For this reason, the trenches can be modified to host the source, while preserving excitation, coupling, and mode propagation (Supplementary Fig. 11).
The dispersion of the ITO layer can be electrostatically tuned by varying the carrier concentration (Fig. 2b), thus changing the behavior at a given operating frequency. We evaluate the spectral response of the metaring, by averaging the field displacement within the air groove for three different carrier concentrations: (i) depletion (ii) ENZ, and (iii) accumulation (Fig. 2c, d).
It is worth observing that for different carrier concentrations we obtain a similar response, energy shifted in the same fashion as at the ENZ resonance, indicating a moderately robust response. Additionally, the spectral response is archetypal of a lowpass frequency filter (longpass wavelength), i.e., integratortype, since we are considering a voltage drop on a capacitor (dielectric air loop) hosted by a resistive circuit board (lossy ENZ) of the equivalent lumped circuit. We then fix the excitation wavelength of the film at the ENZ crossing point for the asdeposited case (1 × 10^{21} cm^{−3}) and take a snapshot of the simulation result of the distribution of the field displacement and current density in the middle plane. In the first scenario (i) when ITO is depleted of its carriers, the material turns dielectric (capacitance, ENZ redshift), providing relatively low losses (\(\varepsilon ^\prime \left( {\lambda _{{{{{{\mathrm{ENZ}}}}}}}} \right)\, > \, 0,\,\varepsilon ^{\prime\prime} \left( {\lambda _{{{{{{\mathrm{ENZ}}}}}}}} \right)\sim 0\)). As a direct consequence, the displacement field in the board hosting the channels becomes nonnegligible; (ii) on the other hand, when the excitation matches the ENZ wavelength (\(\varepsilon ^\prime \left( {\lambda _{{{{{{\mathrm{ENZ}}}}}}}} \right) = 0,\,\varepsilon ^{\prime\prime} \left( {\lambda _{{{{{{\mathrm{ENZ}}}}}}}} \right)\, > \, 0\)), the air loop behaves as Ddot, in which the displacement current circulates in the loop with relatively low losses; (iii) however, when carriers are accumulated, ITO exhibits a predominantly metallic behavior (inductance, ENZ blueshift) characterized by higher ohmic losses (\(\varepsilon ^\prime \left( {\lambda _{{{{{{\mathrm{ENZ}}}}}}}} \right)\, < \, 0,\,\varepsilon^{\prime\prime} \left( {\lambda _{{{{{{\mathrm{ENZ}}}}}}}} \right) \gg 0\)). Equivalent circuit which maps the nanooptic response are displayed as insets (Fig. 2d).
Metaresonator
For demonstrating the generality of this approach, it is possible to introduce reprogrammable nanoelements for implementing different chip functionality. At ENZ wavelength, the specific response of lumped elements (R, L, C) can be achieved by partially filling the air grooves with ITO nanoelements electrostatically modulated with specific voltage polarity. For illustrative purposes, we design the equivalent of an electronic resonator.
Following the characterization of the metaring, next we introduce nanooptic elements within the air trench, thus adding resonator tunability to the system towards adaptation functionality of the nanoscale “filter” (Fig. 3a). The nanoelements are physically realized with the same basic material as the ENZ circuit board, namely ITO, however, bear a different carrier density to exhibit an RL or RC element. Note, this is realized either statically such as via changing the material process conditions during the ITO sputter recipe, and/or actively (i.e., electrostatic tuning) towards delivering elementwise independent tunability of subwavelength small (i.e., lumped) elements (Fig. 3b). As a demonstration, here we consider one element to be injected with carriers (L) and the other one depleted (C), resulting into an LCRlike behavior at the ENZ wavelength. Note, since ITO’s index tunability originates from free carrier modulation and the carrier densities are rather high (even in the depleted case, i.e., depleted here refers to a reduced carrier concentration, not voidofcarriers), an EM wave impeding ITO will always experience a nonzero imaginary optical index, this adding loss, which electrically resembles a resistance R. Thus, a pure C or pure L element is not simply possible in ITO unless gain or losscompensation methods are used. Our model nanocircuit considered here acts as a nanooptic resonator storing energy oscillating at the circuit resonant frequency calibrated to be at ENZ (Supplementary Movie 2).
In an electrical RLC circuit, the voltage drop on a capacitance and inductance are represented by the following rational functions:
The function V_{c}(ω) is characterized by two poles in \(\omega = 1/\sqrt {LC}\), while \(V_L\left( \omega \right)\) has two zeros at 0 frequency and 2 poles in \(\omega = 1/\sqrt {LC}\). Besides the correct mapping of the optical circuit at ENZ, the anticipated lowpass (LP)/highpass (HP) frequency response of the electrical circuit is not perfectly mapped by the dielectric displacement in the ITO nanooptics circuit for each wavelength (Fig. 3d). This is primarily due to three factors:
(i) ITO nanoelements are affected by a strong dispersion are functions of the wavelength unlike the electrical counterpart. More in detail, the optical impedances can be written as:
considering l and t, length and thickness of the nanooptic element, respectively, and ω the optical angular frequency. It can be noticed that the dielectric displacement (and consequently the electric field) within the capacitance is in counter phase with respect to the phase in the inductance (Fig. 3c), as expected. (ii) The staticlike behavior of the metatronic board is guaranteed only at ENZ; (iii) optical losses in the circuit boards.
As shown in Supplementary Movie 2, the energy bounces between a nanooptic inductance and capacitance, thus proving that nanooptic circuits, entirely based on ITO, can be used as programmable oscillators or filters for tuning transmitters and receivers in the telecommunication bandwidth. When electrostatically tuned, the permittivity of the ITO nanoelements can map different problems. Although, when the carriers are electrostatically modulated, Kramers–Kronig (KK) relation dictates that exclusive change of the real part (or imaginary part) of the complex index is not achievable. As the real part decreases with modulation (Δn), the imaginary part increases (Δκ), affecting the performance of the nanooptic circuit, as the equivalent of capacitances or inductances are unwantedly added. Phase change materials that exhibit large Δn with relatively broad transparencies^{38} can be considered for future implementation, especially in regards of their nonvolatile retention of the information, which enables nearzero static power consumption (i.e., no voltage needs to be applied to retain the programmed state). It is also worth mentioning that the ITO region in which the carriers are accumulated is within several nanometers and the 1/e^{2} decaylength is only about 10 nm^{36,39,40}, extended in the simulation to the entire nanoelement. However, the desired extinction ratio can be also adjusted by the applied voltage at the expense of (static) power consumption for a selected design. However, considering the tight confinement within the trenches and slowlight effect, thanks to a high Δn_{eff}, it is possible to achieve high modulation even in relatively short devices^{22}.
Therefore, in principle, the nanooptic elements in the proposed ITObased nonophotonic chip should be deposited using specific process parameters to achieve the desired permittivity according to chosen functionality (capacitance, inductance, and resistance), while finetuning can be achieved using carrier modulation. In terms of PDE mapping, this means that a circuit can solve a class of PDE whose coefficient can be modified by tuning the optical properties of its nanoelement.
Solution of Laplace homogeneous equation using an “ideal” metatronic analog processor
For validating the functionality of the prototyped nanooptic circuit as constituent node of the proposed nanocircuit mesh, we find the solution of a steadystate heat transfer problem in a uniform domain. For demonstration purposes only, we consider the diffusion problem Eq. 1 with Dirichlet boundary conditions, setting the temperature T at the left edge of a rectangular domain L, with H and W height and width of the domain, respectively (Fig. 4a, further details on the derivation provided in the SI). The resistive circuit that solves the finite difference method of a 3 × 3 mesh is shown in Fig. 4b, it can be numerically solved using a “spice tool”, thus obtaining current and voltage distribution. The boundary conditions that mimic the fixed temperature “potential” at the boundary is a voltage generator V, while the heat sink is a simple connection to the ground (relative GND or simply unbiased). The thermal conductivity k of the medium is modeled in the electrical circuit assigning the resistors to be defined as R = L/kA. Similarly, we numerically simulate, by using commercially available fullwave simulation software (further details in Methods section of the SI), the electric field displacement and the displacement current in a 3 × 3 metatronic mesh. A snapshot of the simulation result of the distribution of the electric field in the middle plane is depicted in Fig. 4c. Here, a strong local electric field, generated by a horizontal dipole, is used to model the heat source, while the ENZ condition is applied to the remainder of the boundaries. In this section, as an illustrative and not limiting example, the permittivity of the circuit board is considered to have negligible losses^{41} ε′′ ≈ 0 with an overall size d of 1000 nm, smaller than the operational wavelength (i.e., d < λ) to ensure coupling between the nanoelements, as required for conventional electronic circuit concepts at low frequencies^{16,42}. However, the “spatially staticlike” properties of the ENZ substrate, i.e., absence of a significant phase variation in ENZ (Supplementary Fig. 1 and Supplementary Movie 1), essentially relaxes this requirement for the optical nanocircuit board of Fig. 4c, for which the total length may become also several freespace wavelengths long (while it is electrically small compared to the very long wavelength in ENZ)^{16}. Additional details are provided in Supplementary Note 1.
Under these conditions, the field lines in Fig. 4c highlight that the electric displacement, and consequently the displacement current, fall only within the air grooves, forced by the ENZ conditions in the neighboring area (D ≈ 0). Solution of the PDE problem from different mesh densities in a circuit with ENZ material with negligible losses (ε ≈ 0) suggests mesh scalability (Fig. 4d). Interestingly, solutions of the PDE adopting a metatronic processor with increasing mesh density keeping the overall circuit dimension maps precisely the solution of a finer mesh in a finite difference approach, provided the circuit board is characterized by negligible losses, i.e., absence of a dielectric displacement field in the ENZ circuit board, providing a perfect lumpedelement electriccircuit behavior. However, a study of the size and scalability and their impact on the accuracy of the solution of the metatronic processor becomes determinant if the losses in the ENZ circuit board, or the deviation from the ideal zero permittivity condition, are not negligible. Other parameters, such as width of the waveguide and smoothness of the bending curves, analyzed in Supplementary Fig. 2a, can affect the accuracy of the solution, and will be discussed elsewhere, and systematic errors could be compensated or mitigated by accurate and controlled processes.
Fixing the mesh density of the simulations, the discretized solutions provided by finite differences method (temperature) and the electromagnetic simulation at the node (local electric field displacement) are normalized to their maximum value and compared. The PDE solution shown by this nanophotonic engine is in high qualitative and quantitative agreement (>95%) with respect to solution obtained using commercial software. Further details between the analytic solution of the Laplace equation, the discretized solution obtained through numerical simulations (FDM), and the normalized dielectric displacement at each node for different mesh densities is described in Supplementary Note 3.
Other kind of elliptic PDE can also be solved using a different circuit configuration as described in Supplementary Note 7 where a nonhomogeneous, secondorder, elliptic PDE is solved using a finite difference approach mapped onto a metatronic circuit.
Monolithic integration of reprogrammable metatronic processor for approximate computing
Recently, a few materials have been considered for fabricating a metatronic circuit board, such as multilayered stacks of thin film^{43}, NP assemblies, and graphene^{44}. However, their largescale integration is far from easy. Here we utilize ITO as a suitable material for a monolithic integration of the proposed metatronic processor (Fig. 5a), which enables distinctive functionalities and is being regularly processed by the consumer electronics such as the touchscreen of every tablet and smart phone, for example. The advantages of using ITO are manifolds: (1) ITO has a tunable and controllable optical and electrical properties in the NIR according to process parameters (e.g., oxygen and argon flow rate while sputtering, temperature and environment conditions in thermal annealing processes); (2) as previously demonstrated, its optical properties, imaginary and real part of the permittivity, can be electrostatically tuned^{35,36,45}, by carrier injection, thus allowing potentially GHzfast energyefficient^{22,46} reprogrammability features on the circuit board; and (3) ITO has high manufacturability potential (Supplementary Figs. 6 and 7). Moreover, recently, our group achieved a consistent control over ITO optical parameters with respect to the ENZ wavelength as a function of sputtering parameter, thus allowing to bridge the technological gap in the implementation of metatronic circuits^{21}. According to our experimental studies (additional information on film characterization provided in Supplementary Note 5) depicted in Fig. 5b, the ITO for the ENZ circuit board could potentially be sputtered with 5 standard cubic centimeters per minute (sccm) oxygen flow rate, enabling a 200 nm film in ENZ condition at 1550 nm, with nonnegligible losses \(\widetilde {\upvarepsilon} = 0.3i\) which corresponds to a scattering time ┌ = 0.2 fs. The resistors can be deposited using 20 sccm oxygen flow rate, which yields to \(\widetilde {\upvarepsilon} = 1.2 + 0.6i\) and a scattering time of 5 fs. The main limitations of the ITO metatronic processor are the unwanted losses in the ENZ circuit board.
Because of the losses in the ITO circuit board, the lines of the displacement field are not fully contained in the air grooves, contrarily to the case of an ENZ material with negligible losses, deviating its behavior from a purely lumpedelement circuit. This translates into an inaccurate solution which turns to a faster spatial decay compared to the analytical solution. This deviation is highlighted in Supplementary Fig. 3.
In the presence of nonnegligible losses in the ENZ material, the circuit board is not completely insulating, since the displacement current is not negligible:
There are two major coexisting phenomena that impact the accuracy of the PDE solution, both of which depend on the ITO losses (Fig. 5c): the first one is a function of the mesh density and the second one of the total physical length of the circuit board. High density (>5 × 5) induces coupling within wires that should not be connected, while the larger physical length (>2 µm) contributes to unwanted dissipation, deviating from the original solution. The accuracy as a function of the number of nodes and physical dimension of the circuit board is given in Fig. 5c. The maximum accuracy (>90%) is obtained for a 1 µm size mesh, with a 4 × 4 mesh density. This is achieved thanks to the tradeoff between mesh size and density, which minimizes the wire coupling, without extending the wiring length, producing unwanted losses.
As previously discussed, the ITO layer can be in a capacitor configuration, spaced by a thin dielectric, for electrostatic doping, enabling finetuning of the permittivity values (Fig. 5b) as well as reprogramming the circuit for mapping different problems. The variation of the carrier density via gating in ITO affects both resistance and reactance in the metatronics equivalent circuit, hindering the accuracy of the solution, being imaginary, and the real part of the permittivity. Nevertheless, contrary to the resistive circuit, if either the boundary conditions or the impedances are rapidly “programmed”, the nanooptic equivalent circuit is not affected by dispersion. This is possible because, unlike electric circuit, here the modulation speed can be RCelement large, which is independent from the overall circuit (unlike an electronic circuit). The signal modulation timescales (<ns) are much larger than the propagation delay of the electromagnetic waves across the lumped element. In this case, in fact, the lumped circuit model would hold, since even at 100 s of GHz, the timescale at which the carrier signal is modulated does so substantially slower than the time taken by the optical signal to travel through the nanooptics network.
Additionally, besides affecting the accuracy of the solution, the losses affect the power consumption and dissipation. For a processor with a contained size, an approximate solution is always guaranteed. The power consumption from the processor is the summation of the optical power used for exciting the dipole (initiating the processor) and the radio frequency power employed for modulating the carrier density of each lumped elements, i.e., reprogramming the circuit. Concerning the reconfigurability of the processor, recent works showed few femtojouleefficient ITObased modulators, potentially operating at highspeed^{22,46} (see Supplementary Fig. 7 for GHzfast experimental ITO modulation). On the other hand, a few milliwatts of optical power is needed for exciting the fluorescent molecule and setting the boundary conditions. Efficient measurement schemes must be used to detect the electric field displacement at each node of the metatronic mesh, avoiding scanning over the sample, e.g., highresolution tipenhanced nearfield spectroscopy, to minimize the power used for the detection mechanism.
As a final remark, the constraint on the engine footprint given by the effective wavelength at ENZ and the minimum internode distance, for avoiding unwanted coupling of neighboring wires in a lossy ENZ material, hinders the implementation of high number of nodes. A possible solution would be trading off reconfigurability and using nonreprogrammable nearzero materials characterized by negligible losses or photonic wires^{32}. Other challenges of a possible implementation are discussed in Supplementary Notes 8 and 9.
Discretized solutions with nanooptic probe card
In order to sample the electric field displacement signal at the nodes of the metatronic mesh, deep subwavelength nearfield microscopy must be employed with nanometric spatial resolution, which can be used for investigating the local nearfield, breaking the diffraction limit. (Fig. 6) A sSNOM in transmission, which uses a dielectric tip can be employed for measuring the dielectric displacement, in a similar configuration shown in^{47} with a resolution of a few nanometers in the lateral dimension. The sSNOM would scan the entire ENZ board and its trenches in tapping mode. The radiation is conveyed from the bottom using a focused light which by impinging on a QD, excites a strong nearfield within the trenches. Further details on launching propagating mode using optical antenna or dipole at mesh node are discussed in Supplementary Notes 6 and 9. The scattered radiation from the tip is collected by a parabolic mirror and provides information of the complex optical properties of the nanoscale region in proximity of the tip. By using the sSNOM, it is possible to retrieve both amplitude and phase of the backgroundfree local electric nearfield. However, regular nearfield optical microscopes, such as scatteringtype near field, are associated with AFM systems, thus requiring long scanning time. For this purpose, a nanophotonic probe card could potentially read the values of the local displacement field (see Supplementary Fig. 8), similarly to a wafer probe for electrical testing. The reading mechanism is based on an array of dielectric tips characterized by a subwavelength aperture at the apex^{48}, which collects the local nearfield radiation similarly to a local nearfield microscope, allowing for parallel readout. Aperture SNOM would be preferential with respect to scattering type (sSNOM), since the former will minimize the coupling between vertical dipoles, i.e., metallic tip, while the latter can introduce secondorder scattering and a higher degree of uncertainty in the system.
In conclusion, here we introduced a nanooptic metatronicbased integrated engine, operating in the optical telecommunication band, based on a subwavelength epsilonnear zero (ENZ) circuitry. We mathematically prove that the mesh current method of a metatronic circuit, which mimics a lumpedelement electrical circuit, can map a finite difference mesh that solves a steadystate Laplacian equation and, by extension, other timevariant secondorder elliptic PDEs. Moreover, in contrast with a resistive network, we numerically show that this technology can be used to solve PDEs with high accuracy (90%), while decoupling circuit mesh upscaling from reprogramming speed. Beyond a theoretical framework, we showed through numerical simulation that this metatronic circuit solver can be realized using one single material, exploiting the ability to tailor the optical properties of ITO, sputtered using, experimentally validated, controlled process parameters, thus enabling a high degree of manufacturability. This analog processor is particularly wellsuited to be reprogrammed on both the boundary conditions and network elements, limited in size only by the presence of ohmic losses. Implementing these techniques enables an ultrafast, chipscale, integrable, and reconfigurable analog computing processor that is able to solve PDEs at the speed of light without lumped circuit capacitance delay.
Methods
The numerical simulations of this study were carried out using commercially available software Comsol Multiphysics, we used the frequencydependent solver with a tetrahedron mesh. The dimension of the maximum meshing element of the metatronic board was 1/100 of λ and a denser local mesh was considered for the air grooves and the lumped components. The electric field was excited by a point dipole moment of amplitude 1 V. The ITO film’s optical properties (complex permittivity and thickness) and relative sputtering process parameters (oxygen/argon flow rate, deposition time, annealing time and temperature) were taken from Silva et al.^{18}. The effects of the carrier injection to the dispersion of the ITO film in the nearinfrared region was derived using Drude’s model. Complex refractive index, scattering rate, ENZ wavelength, initial carrier concentration, thickness, and plasma frequency were derived by spectroscopic ellipsometry of the deposited film and used as fitting parameters for the active tuning. Further details regarding the methods used in this study can be found in Supplementary Note 5.
Data availability
All data needed to evaluate the conclusions in the paper are present in the paper or the Supplementary Online Materials. Additional data available from authors upon request.
References
Tsividis, Y. Not your Father’s analog computer.IEEE Spectrum 55, 38–43 (2018).
Ambrogio, S. et al. Equivalentaccuracy accelerated neuralnetwork training using analogue memory. Nature 558, 60 (2018).
Zidan, M. A. et al. A general memristorbased partial differential equation solver. Nat. Electron. 1, 411–420 (2018).
Shen, Y. et al. Deep learning with coherent nanophotonic circuits. Nat. Photonics 11, 441–446 (2017).
Miscuglio, M. et al. Alloptical nonlinear activation function for photonic neural networks [Invited]. Opt. Mater. Express, OME 8, 3851–3863 (2018).
Ríos, C. et al. Inmemory computing on a photonic platform. Sci. Adv. 5, eaau5759 (2019).
Miscuglio, M. & Sorger, V. J. Photonic tensor cores for machine learning. Appl. Phys. Rev. 7, 031404 (2020).
Lin, X. et al. Alloptical machine learning using diffractive deep neural networks. Science 361, 1004–1008 (2018).
Mohammadi Estakhri, N., Edwards, B. & Engheta, N. Inversedesigned metastructures that solve equations. Science 363, 1333–1338 (2019).
Heuck, M., Jacobs, K. & Englund, D. R. Controlledphase gate using dynamically coupled cavities and optical nonlinearities. Phys. Rev. Lett. 124, 160501 (2020).
Liebmann, G. Solution of partial differential equations with a resistance network analogue. Br. J. Appl. Phys. 1, 92–103 (1950).
Rocca, E. & Rossi, R. ‘Entropic’ solutions to a thermodynamically consistent PDE system for phase transitions and damage. SIAM J. Math. Anal. 47, 2519–2586 (2015).
Borggaard, J. & Burns, J. A PDE sensitivity equation method for optimal aerodynamic design. J. Comput. Phys. 136, 366–384 (1997).
Crank, J. & Nicolson, P. A practical method for numerical evaluation of solutions of partial differential equations of the heatconduction type. Adv. Comput. Math. 6, 207–226 (1996).
Chesshire, G. & Henshaw, W. D. Composite overlapping meshes for the solution of partial differential equations. J. Comput. Phys. 90, 1–64 (1990).
Li, Y. et al. Onchip zeroindex metamaterials. Nat. Photonics 9, 738–742 (2015).
Caglayan, H., Hong, S.H., Edwards, B., Kagan, C. R. & Engheta, N. Nearinfrared metatronic nanocircuits by design. Phys. Rev. Lett. 111, 073904 (2013).
Silva, A. et al. Performing mathematical operations with metamaterials. Science 343, 160–163 (2014).
Lumer, Y., Liberal, I. & Engheta, N. Quantum features of optical metatronics. In 2017 Conference on Lasers and ElectroOptics (CLEO) 1–2 (OSA, 2017).
Lumer, Y. & Engheta, N. Circuit modularization of quantum optical systems. Phys. Rev. Appl. 14, 054034 (2020).
Gui, Y. et al. Towards integrated metatronics: a holistic approach on precise optical and electrical properties of indium tin oxide. Sci. Rep. 9, 11279 (2019).
Amin, R. et al. Subwavelength GHzfast broadband ITO Mach–Zehnder modulator on silicon photonics. Optica 7, 333–335 (2020).
Chaudhari, P., Oberman, A., Osher, S., Soatto, S. & Carlier, G. Deep relaxation: partial differential equations for optimizing deep neural networks. Res. Math. Sci. 5, 30 (2018).
Perona, P. & Malik, J. Scalespace and edge detection using anisotropic diffusion. IEEE Trans. Pattern Anal. Mach. Intell. 12, 629–639 (1990).
Aletti, G., Moroni, M. & Naldi, G. A new nonlocal nonlinear diffusion equation for data analysis. Acta Appl. Math. 168, 109–135 (2020).
Palmer, P. J. Investigations Into the Use of an Electrical Resistance Analogue for the Solution of Certain Oscillatoryflow Problems (H.M. Stationery Office, 1957).
Liebmann, G. Resistancenetwork analogues with unequal meshes or subdivided meshes. Br. J. Appl. Phys. 5, 362–366 (1954).
Edwards, B. & Engheta, N. Experimental verification of displacementcurrent conduits in metamaterialsinspired optical circuitry. Phys. Rev. Lett. 108, 193902 (2012).
Liu, N. et al. Individual nanoantennas loaded with threedimensional optical nanocircuits. Nano Lett. 13, 142–147 (2013).
Shi, J. et al. Modular assembly of optical nanocircuits. Nat. Commun. 5, 1–8 (2014).
Alù, A. & Engheta, N. All optical metamaterial circuit board at the nanoscale.Phys. Rev. Lett. 103, 143902 (2009).
Liu, R., Roberts, C. M., Zhong, Y., Podolskiy, V. A. & Wasserman, D. Epsilonnearzero photonics wires. ACS Photon. 3, 1045–1052 (2016).
Liberal, I., Mahmoud, A. M., Li, Y., Edwards, B. & Engheta, N. Photonic doping of epsilonnearzero media. Science 355, 1058–1062 (2017).
Boltasseva, A. & Atwater, H. A. Lowloss plasmonic metamaterials. Science 331, 290–291 (2011).
Lee, H. W. et al. Nanoscale conducting oxide PlasMOStor. Nano Lett. 14, 6463–6468 (2014).
Sorger, V.J., LanzillottiKimura, N.D., Ma, R.M. & Zhang, X. Ultracompact silicon nanophotonic modulator with broadband response. Nanophotonics 1, 17–22 (2012).
Lu, H., Carroll, G. M., Neale, N. R. & Beard, M. C. Infrared quantum dots: progress, challenges, and opportunities. ACS Nano. 13, 939–953 (2019).
Zhang, Y. et al. Broadband transparent optical phase change materials for highperformance nonvolatile photonics. Nat. Commun. 10, 1–9 (2019).
Feigenbaum, E., Diest, K. & Atwater, H. A. Unityorder index change in transparent conducting oxides at visible frequencies. Nano Lett. 10, 2111–2116 (2010).
Li, E., Gao, Q., Chen, R. T. & Wang, A. X. Ultracompact siliconconductive oxide nanocavity modulator with 0.02 lambdacubic active volume. Nano Lett. 18, 1075–1081 (2018).
Silveirinha, M. G., Alù, A., Li, J. & Engheta, N. Nanoinsulators and nanoconnectors for optical nanocircuits. J. Appl. Phys. 103, 064305 (2008).
Moitra, P., Slovick, B. A., Gang Yu, Z., Krishnamurthy, S. & Valentine, J. Experimental demonstration of a broadband alldielectric metamaterial perfect reflector. Appl. Phys. Lett. 104, 171102 (2014).
Li, Y., Liberal, I., Della Giovampaola, C. & Engheta, N. Waveguide metatronics: lumped circuitry based on structural dispersion. Sci. Adv. 2, e1501790 (2016).
Vakil, A. & Engheta, N. Transformation optics using graphene. Science 332, 1291–1294 (2011).
Tahersima, M. H. et al. Couplingenhanced dual ITO layer electroabsorption modulator in silicon photonics. Nanophotonics 8, 1559–1566 (2019).
Li, E., Nia, B. A., Zhou, B. & Wang, A. X. Transparent conductive oxidegated silicon microring with extreme resonance wavelength tunability. Photon. Res. 7, 473–477 (2019).
AlonsoGonzález, P. et al. Controlling graphene plasmons with resonant metal antennas and spatial conductivity patterns. Science 344, 1369–1373 (2014).
Bao, W. et al. Mapping local charge recombination heterogeneity by multidimensional nanospectroscopic imaging. Science 338, 1317–1321 (2012).
Acknowledgements
This work was supported by National Science Foundation (1748294) grant.
Author information
Authors and Affiliations
Contributions
V.J.S. and T.E.G. envisioned the idea of a reprogrammable optical computer, acquired the funds, and supervised the project. M.M. and V.J.S. conceived the nanooptic platform for analog computing. M.M. developed the relevant theories and analyses for the project, and designed the nanooptic circuit. Y.G. and M.M. conducted the spectroscopy ellipsometry experiments. Z.M., S.S., and X.M. conducted numerical simulations. M.M., A.A., and V.J.S. discussed the results and contributed to the understanding, analysis, and interpretation of the results. M.M. wrote the first draft of the manuscript. V.J.S., A.A., T.I., and T.E.G. contributed to writing subsequent drafts of the manuscript.
Corresponding authors
Ethics declarations
Competing interests
The authors have filed a provisional patent application on this idea, U.S. Patent 10,318,680. No other conflict of interest is present.
Additional information
Peer review information Communications Physics thanks the anonymous reviewers for their contribution to the peer review of this work.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Miscuglio, M., Gui, Y., Ma, X. et al. Approximate analog computing with metatronic circuits. Commun Phys 4, 196 (2021). https://doi.org/10.1038/s42005021006834
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s42005021006834
This article is cited by

Modeling and analysis of a photonic crystal embedded ENZ gyrotropic metatronic amplifier using the mode matching technique
Optical and Quantum Electronics (2024)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.