Non-volatile materials for programmable photonics

Programmable photonics play a crucial role in many emerging applications, from optical accelerators for machine learning to quantum information technologies. Conventionally, photonic systems are tuned by mechanisms such as the thermo-optic effect, free carrier dispersion, the electro-optic effect, or micro-mechanical movement. Although these physical effects allow either fast ( > 100 GHz) or large contrast ( > 60 dB) switching, their high static power consumption is not optimal for programmability, which requires only infrequent switching and has a long static time. Non-volatile materials, such as phase-change materials, ferroelectrics, vanadium dioxide, and memristive metal oxide materials, can offer an ideal solution thanks to their reversible switching and non-volatile behavior, enabling a truly “set-and-forget” programmable unit with no static power consumption. In recent years, we have indeed witnessed the fast adoption of non-volatile materials in programmable photonic systems, including photonic integrated circuits and free-space meta-optics. Here, we review the recent progress in the field of programmable photonics, based on non-volatile materials. We first discuss the material’s properties, operating mechanisms, and then their potential applications in programmable photonics. Finally, we provide an outlook for future research directions. The review serves as a reference for choosing the ideal material system to realize non-volatile operation for various photonic applications. © 2023 Author(s). All article content, except where otherwise noted, is licensed under a Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/). https://doi.org/10.1063/5.0165309


INTRODUCTION
Programmability is an increasingly important feature in modern-day photonic systems and is crucial to enabling technologies ranging from photonic accelerators for machine learning 1 to quantum information processing. 2 Such programmability is generally achieved by traditional modulation methods, including thermo-optic (TO), 3 free-carrier dispersion, 4 electro-optic (EO), 5 or mechanical tuning. 6These physical effects either incur a large static power consumption (>10 mW in the case of the TO effect) or require a constant bias (in the case of the EO effect and mechanical tuning) when the photonic systems are in the standby state.Since the programmability only requires infrequent switching (normally <1 kHz) and the photonic systems stay in the standby state most of the time, the static energy consumption can easily scale with the number of photonic switches and tremendously increase the power budget of the entire photonic system.The second major challenge faced by programmable photonics is that fabrication variation inevitably results in phase errors in the systems, particularly for high-index-contrast platforms such as silicon-on-insulators. 7hile wider or thicker waveguides and Si 3 N 4 platforms have been proposed to improve the fabrication tolerance, 7 the most effective solution so far has been the use of TO phase shifters to correct phase errors in the resonators or interferometers.This will further incur large static power consumption and thermal crosstalk since these phase shifters are always in the on state.Therefore, by reducing or even removing such static power consumption, we can drastically improve the energy efficiency and reduce the crosstalk of any REVIEW pubs.aip.org/aip/apmprogrammable photonic systems, including routing light in optical interconnects, 8 setting phase masks in spatial light modulators (SLMs), 9 semi-static displays, 10 or trimming photonic resonators to the same resonance frequency. 11,12In fact, this can be readily achieved by using non-volatile tuning mechanisms where the optical state can be maintained without any power consumption or bias, akin to a "set-and-forget" switch.Non-volatile materials, such as phase-change materials (PCMs), memristive materials, ferroelectric materials, and ferromagnetic materials, are attractive candidates to realize non-volatile tuning as they have proven to be reliable memory mediums in electronics and optical storage.For example, PCMs are the enabling technology behind rewritable compact disks 13 and phase-change random access memory (RAM); 14 Memristive materials are the building blocks of Resistive RAM (ReRAM); 15 Ferroelectric RAM (FeRAM) and ferroelectric field effect transistors (FeFETs) are based on ferroelectric materials; 16 and ferromagnetic materials are the key elements for magnetoresistive RAM (MRAM). 17One exception is flash memory, which involves only CMOS compatible materials, normally thin tunneling dielectrics and a charge-trapping silicon layer, to achieve the desired memory effect. 18Figure 1 shows the timeline of the development of various non-volatile material platforms for programmable photonics and memory applications.A REVIEW pubs.aip.org/aip/apmsilicon nanocrystal optical memory was first demonstrated in 2004 based on the charge-trapping effect, 19 which is read optically via the nanocrystals' photoluminescence into free space.Photonic integration of charge-trapping flash memory onto silicon waveguides was first theoretically proposed 20 in 2006 and demonstrated experimentally ten years later. 21On the other hand, PCMs were first proposed to be used for optical switching in a silicon waveguide in 2010 22 and later in integrated photonic memory 23 and optical computing. 24A photonic memristor based on amorphous silicon was first demonstrated in 2013. 258][29][30][31] Active photon sources were discovered in an Ag/amorphous-SiO/Pt plasmonic memristor. 32A non-volatile magneto-optic silicon photonic phase shifter based on ferromagnetic CoFeB was reported in 2020. 33Ferroelectric materials such as Barium Titanate (BTO) have recently enabled a multilevel nonvolatile silicon photonic phase shifter. 34Finally, vanadium dioxide (VO 2 ), a material previously considered to be volatile, was recently discovered to exhibit a memory effect under a constant voltage bias. 35Due to the rising interest in this field, in this article, we review the recent progress in the applications of non-volatile materials in programmable photonics and optical memory.We first explain the underlying mechanisms responsible for the non-volatile effect in these materials and then compare their key performance metrics, such as switching speed, endurance, and switching energy, before discussing their suitability for various photonic applications.Several future research directions are also identified.The article aims to serve as a general guideline for researchers who are looking for the most suitable non-volatile material systems for their desired photonic applications.7][38][39] The latching effect of micro-electromechanical systems (MEMS) can also enable non-volatile photonic switching 40 but is out of the current scope, which focuses mainly on the non-volatile property of the photonic material.

PHASE-CHANGE MATERIALS AND PHASE-CHANGE MEMORY
Phase-change materials (PCMs), exemplified by Ge 2 Sb 2 Te5 (GST), are chalcogenides that can be switched between bistable micro-structural states-crystalline and amorphous states.Figure 2 shows that to turn the materials from an ordered crystalline state to a disordered amorphous state, we apply a high power and short excitation to the PCMs that essentially melt-quench the materials into the amorphous state.The high-power excitation raises the material temperature above the melting point, and the short duration ensures it is cooled rapidly after the excitation is removed.To return the material to the crystalline state, a moderate amplitude and long excitation a applied that anneal the material into the crystalline state.The moderate power raises the material temperature above the crystallization temperature and below the melting point, and the long duration facilitates the long-range diffusion of atoms to settle into their most thermodynamically favorable crystalline state.Such thermal excitation can be achieved either via optical or electrical pulses.Since crystalline PCM has a higher refractive index than its amorphous FIG. 2. Operating principle of phase-change materials in programmable photonics.PCMs tend to have a higher refractive index in the crystalline state due to the meta-valent bonding, 50 and a lower index in the amorphous state due to the covalent bonding.A high amplitude and short excitation are applied to switch the PCM into the disordered amorphous state.The excitation melts the material before quenching it rapidly to prevent long-range diffusion.To return the material to its crystalline state, a moderate amplitude and long excitation are used to anneal the material into an ordered crystalline state.For programmable photonics, typically ∼100 mW of power and ∼100 ns pulse width are required for RESET. 51For SET, the power is ∼10 mW and the pulse width is around tens of μs.However, the switching power and time do scale with the volume of the PCMs and can go down to only ∼μW and ∼ns in nanoscale electronic memory. 52n (refractive index); k (extinction coefficient); Tm (melting point); Tx (crystallization temperature).
state due to the increased electron delocalization, 41 we can harness such a large refractive index contrast (typically Δn >1) to control the phase and absorption of the light propagating inside the material. 42n the early days, PCMs were widely used in rewritable compact disks to store data.They were typically designed to be highly absorptive at the visible wavelength so that they could be easily switched by laser pulses.Later, they found applications in electronic memory, where data are stored in the high and low resistance of the memory cell, rendering the optical absorption of the materials entirely irrelevant.It is only recently, with the advent of PCM-integrated photonics, 43,44 that absorptive loss has become critical.The proximity of the PCMs to the optical mode in a waveguide can lead to a significant loss even for a very short PCM segment, e.g., 10 nm crystalline GST typically gives >1 dB/μm insertion loss on a Si 3 N 4 waveguide. 45To address this issue, new PCMs have been discovered that have wider bandgaps and hence lower losses than GST, 46 such as GSST, 47 GSSe, 47 Sb 2 Se 3 , 48 and Sb 2 S 3 . 48,49

APPLICATIONS OF PHASE-CHANGE MATERIALS IN PHOTONICS
We categorize the applications of PCMs into four quadrants depending on whether they are intended for free space or photonic integrated circuits (PICs) and whether the control is via electrical or optical means (see Fig. 3).The earliest work can be traced back to Tsuda, who first used a laser to switch Ge 2 Sb 2 Te5 (GST) on a silicon waveguide and demonstrated an ultra-compact optical switch. 22he same group improved on their previous design to show a selfholding optical switch based on a multi-mode Si waveguide that can be reversibly switched over 2000 cycles. 53Rudé et al. first showed the control of a microring by laser switching GST 54  Beyond PICs, PCMs also find applications in free space.Early work envisioned an electrically tunable reflective display based on GST where each pixel can be individually switched using ITO electrodes 10 [Fig.3(c-i)].Such a rewritable color nanoprint has also been achieved in wide bandgap PCMs Sb 2 S 3 controlled by a femtosecond laser 68 [Fig.3(d-iv)].Wang et al. first demonstrated an optically reconfigurable metasurface where they showed phase masks of zone plates and gratings, etc. can be written into a blanket thin film of GST 65 [Fig.3(d-i)].Using Huygen's metasurface, near 2π phase control was realized 66 [Fig.3(d-iii)].A tunable spectral filter in the mid-IR was demonstrated using a GST-Ag plasmonic metasurface. 67imilar functionalities such as spectral tuning and beam steering were also demonstrated using metasurfaces via electrical control by metallic microheaters in visible and near IR [62][63][64] [Figs.3(c-ii)-3(c-iv)].For a more detailed discussion, please also refer to other review papers on PCM-based photonics. 43,44,69

MEMRISTORS AND RESISTIVE MEMORY
A memristor consists of a metal-dielectric-metal or semiconductor-dielectric-semiconductor tri-layer (see Fig. 4).Typical choices of metals include Pt, Ti, Ta, andAl, and the dielectrics are oxides such as TiO 2 , NiO, Al 2 O 3 , HfO 2 , TaO 2 , etc.Note that the top and bottom metal or semiconductor layers do not have to be identical and are typically chosen to aid the process of catalyzing anions or cations in the memristors. 70At moderate bias across the junction, the material stack acts as a capacitor.Beyond a certain threshold voltage, the ions inside the dielectric, for example, oxygen ions inside transitional metal oxides, 71,72 start to migrate toward the anode.In some situations, metal ions from the electrodes, for example, Ta cations from the Ta electrodes, 73 start to migrate toward the cathode.If we consider the simple case of oxygen ion migration, then simultaneously, the oxygen vacancies migrate toward the cathode.A conduction filament (CF) is formed by the percolation of oxygen vacancies under a high electric field, and its growth is promoted by the subsequent Joule heating. 74The first formation of the CFs in a pristine dielectric is called "electroforming," which brings the memristor to the low resistance state (LRS).It generates defects inside the dielectric and makes subsequent switching of the memristor possible.Once the CF forms, the memristor can be reset by reversing the bias polarity in the case of bipolar switching or using the same polarity in the case of unipolar switching.The RESET causes the rupture of the CFs due to a combined effect of Joule heating and an electric field. 75For bipolar switching (depicted in Fig. 4), it is believed that the electric field plays a more significant role in CF rupture, while for unipolar switching, Joule heating seems to play a more dominant role. 75The breaking of CFs brings the memristor to the high resistance state (HRS).Instead of dissolving the entire CF, the RESET normally causes the breaking of CF only at the thinnest section, where the resistance is the largest. 72Hence, to set the memristor back to the LRS, a lower voltage is required than the electroforming to reconnect the CF.The memristor can be FIG.4. Operating principle of a memristor under bipolar switching.A memristor typically consists of a conductor-insulator-conductor tri-layer.The conductors can be either metals or semiconductors while the insulator is generally dielectric materials.A conduction filament is formed by applying a large enough voltage which switches the memristor from the initial HRS into the LRS.This process is called "electroforming."The memristor can be returned to the HRS by reversing the bias polarity which breaks the filaments.The switching is reversible and non-volatile as the filament remains after the bias is removed.Since the conduction filaments do not break completely during RESET, the SET normally require lower voltage than the electroforming.LRS (HRS), low resistance state (high resistance state).

REVIEW
pubs.aip.org/aip/apmreversibly switched between HRS and LRS for millions of cycles, 76 and the excitation can either be direct current (e.g., an IV sweep) or pulses.So far, two approaches have been studied to couple the memristive switching in the electronic domain to the optical domain.8][79] This approach leads to large optical contrast (>10 dB) without the need for a read bias, but the insertion loss is high.The second approach is to combine memristive switching with the free carrier dispersion effect. 30,31Under a small read bias, the free carriers will accumulate at the metal-oxide-semiconductor capacitor when the memristor is in the HRS and leak through the dielectric when the memristor is in the LRS.The advantage of this approach is that the plasmonic waveguides are no longer needed, and loss can be reduced, but a bias will be required to read the optical states.

FIG. 5. Applications of memristors in programmable photonics. (a)
A Au/SiO 2 /ITO plasmonic memristor. 77The plots show a clear hysteresis in optical transmission upon switching the plasmonic memristor on and off by sweeping the voltage.Extinction ratios of 12 dB are demonstrated in 10 μm devices at operating voltages of ±2V.(b) Single-atom switching in a Ag/aSi/Pt plasmonic memristor. 78Switching of the plasmonic resonance in the 1550 nm wavelength range is shown, leading to a large extinction ratio of 10 dB.(c) Silicon photonic memory based on an GaAs/Al 2 O 3 /Si memristor integrated with a microring resonator. 30The TEM shows the heterogeneous integration of III-V materials with Si and a 20 nm bonding oxide layer in between.(d) Silicon photonic phase shifter based on an n-InP/HfO 2 /p-Si memristor. 31The I-V curve shows typical memristor switching in the electrical domain, while the transmission spectrum shows the switching of microring resonance in the optical domain.

REVIEW pubs.aip.org/aip/apm
Such bias can be applied via ultra-short read pulses to reduce static power consumption. 80Moreover, in the LRS state, the DC bias would lead to static power due to the finite leakage current.These two approaches will be discussed in detail in the section below.

APPLICATIONS OF MEMRISTORS IN PHOTONICS
A recent review paper has discussed the recent progress in various memristor technologies for photonic applications. 81Here, we approach the same topic differently by separating the field into two main research directions, i.e., plasmonic and photonic memristors.Emboras et al. first proposed a plasmonic memristor in an Ag-Si plasmonic waveguide. 25The memristive switching is caused by the formation and breaking of Ag filaments inside the amorphous Si dielectric [see Fig. 1(b)].The subwavelength confinement of the plasmonic waveguide ensures maximal overlap between the mode and the nanoscale filaments.Based on a similar design logic, Fig. 5(a) shows an Au/SiO 2 /ITO plasmonic memristor demonstrated by Hoessbacher et al. with 39% of the field confined in the SiO 2 layer, resulting in a large extinction ratio of 12 dB.It was further shown that only a few atoms are enough to trigger large extinction ratio switching (9.2 dB) in an Ag/Si/Pt plasmonic memristor 78 [see Fig. 5(b)].Apart from optical switching, the plasmonic memristor can also be used for photodetection, where the atomic-scale filaments can be ruptured by a small increase in light intensity. 82n addition, by evanescently coupling light from a Si waveguide to a memristor, the linearity of the memristive switching can be enhanced by a factor of four compared to the absence of light radiation due to the increased oxygen vacancies and diffusion rate. 83he near analog operation is shown to be beneficial for neuromorphic computing. 83A recent study showed that memristors can act as photon source-the formation and dissolution of CFs will lead to the emission of photons-which can potentially be used as single photon emitters for quantum applications or light sources for optical communications. 32Finally, it was discovered that the memristor can also tune the plasmonic scattering of visible light (∼600 nm) from an Ag nanodot on an Au mirror. 84This phenomenon can be used to uncover the morphological changes of the CFs during switching in a non-destructive way.Although plasmonic memristors can achieve large optical contrast by switching atomic scale nanofilaments, the high insertion loss (∼10 dB for a 10 μm long device) can be prohibitive for large-scale PICs.More recently, a silicon photonic memristor based on heterogeneous integration of III-V materials onto a silicon waveguide was demonstrated with low insertion loss for the first time by Tossoun et al. (∼0.05 dB for a 10 μm radius microring). 30The memristor is formed by an n-type GaAs and a p-type Si sandwiching a thin layer of Al 2 O 3 inside a microring resonator.Since the mode is largely confined inside the silicon [Fig.5(c)], the loss is significantly reduced, but the formation of nano-filament does not couple very well to the optical domain due to the small mode overlap.To enhance the effect of memristive switching, a bias is normally applied across the FIG. 6.Operating principle of flash memory.A flash memory cell is based on a traditional MOSFET configuration with an additional floating gate, typically made of polysilicon.Depending on the polarity of the gate voltage, the electrons can be shifted back and forth between the floating gate and the conduction channel via tunneling.The carriers are trapped inside the floating gate by the thin dielectric underneath it after the bias is removed, which provides the memory effect.This can be readily converted into a photonic memory by etching the p-type silicon into a waveguide.Light propagating within the waveguide will interact with the trapped charges via free carrier dispersion: strong dispersion when there are trapped charges and weak dispersion when there are no charges.

REVIEW
pubs.aip.org/aip/apmmetal-oxide-semiconductor capacitor to optically read the resonance shift.The memristor switches the operating mode from capacitive carrier accumulation (spectral blue shift) to leakage-resulted carrier injection (spectral red shift), resulting in a difference in the resonance wavelength.Since the switching occurs at the nanoscale filament, the photonic memristor can be switched at high speed (300 ps) and consumes ultra-low energy (0.15 pJ).Based on a similar configuration, a non-volatile phase shifter has also been demonstrated with an n-type InP/HfO 2 /p-type Si memristor 31,85 [see Fig. 5(d)].The memristor is electrically written using pulses and optically read with a −3V bias, which results in a non-volatile shift in the microring resonance between the LRS and HRS.The non-volatile phase shifter has a low switching energy of ∼400 fJ with a Lπ of ∼0.5 mm.

FLASH MEMORY
Similar to memristors, flash memory also leverages the capacitive effect for switching.Flash memory is a non-volatile memory based on a floating gate MOSFET 18 (see Fig. 6).Compared to conventional MOSFETs, a floating gate (FG), which typically consists of polycrystalline Si, is inserted into the dielectric layer, typically made up of SiO 2 , underneath the control gate (CG).Applying an ON-voltage to the CG will allow electrons to flow from the source FIG. 7. Applications of flash memory in programmable photonics.(a) A silicon photonic read-only memory based on a Si 3 N 4 floating gate. 20A microring-based photonic memory was first proposed in simulation, which shows that optical information can be stored in different microring resonance states, controlled by the number of trapped charges.(b) Experimental demonstration of a multilevel programmable silicon photonic memory based on a poly-Si floating gate. 21The resonance wavelengths of the on and off states were shown to be stable over 20 hours, and the device can be reversibly switched for 30 cycles.(c) An optically written silicon photonic flash-memory. 86UV light is used to optically excite the free carriers into the Si 3 N 4 floating gates.(d) A silicon photonic flash memory with a graphene floating gate. 87Graphene can potentially give a larger effective index change per unit of trapped charges compared to poly-Si.(e) A silicon photonic memory with an ITO floating gate. 88ITO can be doped by the trapped carriers into the epsilon-near-zero regime, where a large extinction ratio over 10 dB can be realized for a 5 μm long electro-absorption switch.(f) A III-V-on-Si photonic flash memory based on a Mach-Zehnder interferometer. 89The dielectric layer consists of alternating layers of Al

REVIEW
pubs.aip.org/aip/apm to the drain.At high source-drain current, electrons gain enough energy to overcome the oxide-silicon energy barrier, and the channel hot electrons are injected into the FG at the drain side.The further electron injection into the top control gate is stopped by the thick blocking dielectric, normally formed by a triple layer of oxide-nitride-oxide.When the bias is removed, the electrons in the floating gate are trapped there.This is called the "Write" operation.
To remove the trapped electrons from the FG, the bias polarity at the CG is reversed, forcing the electrons to tunnel back to the ptype Si via Fowler-Nordheim tunneling.This is called the "Erase" operation.If the p-type Si is etched into a waveguide, same configuration can be used for a non-volatile photonic flash memory.
The trapping of electrons inside the FG will modify the effective index of the Si waveguide via free carrier dispersion in a non-volatile way.The concept was first proposed by Barrios and Lipson, where a microring resonator is used to achieve a large modulation depth (91%) 20 [see Fig. 7(a)].

APPLICATIONS OF FLASH MEMORY IN PROGRAMMABLE PHOTONICS
The first silicon photonic flash memory consists of a poly-Si control gate and a Si 3 N 4 floating gate embedded in a SiO 2 matrix.SiO 2 acts both as a blocking and tunneling dielectric [Fig.1(a)]. 20he NPN junction is defined in the partially etched Si waveguide.The experimental demonstration of a silicon photonic flash memory was first reported by Song et al., where they achieved reversible and non-volatile switching of microring resonance 21,90 [see Figs.1(c) and 7(b)].The change in both the resonance wavelength and the resonance Q factor indicates that non-volatile switching is indeed caused by the free carrier dispersion of the charges stored in the FG.Multilevel operation is possible by varying the pulse width from 0 to 625 ms with a maximum resonance shift of 380 pm.Finally, a 4-bit memory was demonstrated by controlling four microrings of different radii coupling to a bus waveguide and reading at four different wavelengths.A slightly different configuration was demonstrated by Grajower et al. with a Si 3 N 4 FG 86 [Fig.7(c)].Instead of electrically writing the memory cell, the carriers are injected into the FG via optical excitation (UV illumination), leading to a resonance shift of 125 pm.The thick poly-Si or Si 3 N 4 used for FG tends to have minor overlap with the optical mode, which leads to a small effective index change.It is shown that graphene can give a larger effective index change per unit of trapped charges compared to poly-Si 87 [Fig.7(d)].Depending on the stored charge density, the optical switch can operate either as an intensity switch or a phase shifter with low insertion loss in the Pauli blocking regime. 91ITO can also be used for the FG due to its low loss and high conductivity, 88 but in reality, the loss of ITO does vary significantly with its carrier concentration [Fig.7(e)].Notably, ITO can be doped by the trapped carriers into the epsilon-near-zero regime, where a large extinction ratio over 10 dB can be realized for a 5 μm long electro-absorption switch.To improve the modulation depth, one can also design the waveguide such that the light is mainly confined in the poly-Si and the FG, which maximizes the overlap between the optical field maxima and the trapped charges. 92HfO 2 is used as the FG due to its high trap density.A III-V-on-Si photonic flash memory was reported recently in a Mach-Zehnder interferometer, which can be switched reversibly for 100 cycles at <20 pW power consumption. 89

FERROELECTRIC MATERIALS AND FERROELECTRIC FIELD EFFECT TRANSISTOR
Ferroelectric materials exhibit nonzero spontaneous polarization even when the applied electric field is zero.The polarization can also be reversed by applying a strong enough electric field in the opposite direction.The ability to maintain polarization after the electric field is removed makes them ideal for memory applications.The first application of ferroelectric materials for storage is in FeRAM using lead zirconate titanate (PZT) or strontium bismuth tantalate (SBT), later commercialized by companies such as Samsung and Sony. 93Traditional FeRAM cell consists of a ferroelectric capacitor connected to the drain of a MOSFET. 16,93Connecting the capacitor in series with the transistor means that reading the cell will inevitably lead to bit erasure.The discovery of ferroelectricity in Si-doped HfO 2 in 2011 has revived interest in FeFET to realize a highly scalable non-volatile memory. 94This is because HfO 2 is a more CMOS compatible material, which has been widely used as a gate dielectric, than the traditional perovskite-based materials such as PZT, which are challenging to integrate with the CMOS process.Additionally, since the ferroelectric material now acts as the gate dielectric, the read-out can be non-destructive.The MOSFETlike configuration means the FeFET has both memory and logic functions, so it can be used for in-memory computing or artificial neurons for spiking neural networks. 94Figure 8 shows the operating principle of a FeFET.The interfacial layer is a thin dielectric, such as SiO 2 or Si 3 N 4 , that prevents charge leakage and the diffusion of Si into the ferroelectric, which can degrade its properties. 95he FeFET is written into the ON state (polarization down) when the gate voltage is larger than the positive coercive voltage Vc, which is the bias required to eliminate the macroscopic polarization of the ferroelectrics.This will lead to the accumulation of electrons underneath the gate dielectric, and the drain-source current will increase.On the other hand, the FeFET can be erased by applying an opposite gate voltage that is larger than the negative Vc.This brings the FeFET into the OFF state (polarization up), and electrons in the channel are depleted, causing the drain-source current to drop.As the switching is solely a field driven effect from a low program voltage (typically <5V), FeFET is expected to be a highly energy efficient memory.

APPLICATIONS OF FERROELECTRIC MEMORY IN PROGRAMMABLE PHOTONICS
Although ferroelectric materials have long been explored in photonics, their applications were limited to high-speed modulators, which leveraged the strong Pockels effect of ferroelectric materials such as PZT 96 and Barium Titanate (BTO). 97The non-volatile effect of ferroelectric in photonics has not been studied until recently, when a silicon photonic phase shifter based on BTO was reported 34 (see Fig. 9).Like a FeFET, the polarization of the ferroelectric BTO layer underneath the silicon can be switched using electrical pulses of different amplitudes or pulse widths, and the bias polarity will determine the polarization orientation.Depending on the orientation of the ferroelectric domains, the Pockel's coefficient changes in a nonvolatile fashion, which in turn leads to a refractive index change if the phase shifter is read with a fixed DC voltage bias.Since the ferroelectric switching is entirely based on a field effect, the switching energy can be very low (4.6 pJ).Meanwhile, the use of Pockel's effect REVIEW pubs.aip.org/aip/apmFIG. 8. Operating principle of a FeFET.A FeFET has a similar structure to a MOSFET, where the gate dielectric is replaced with a ferroelectric material.Switching is achieved by applying a bias to the gate, which induces non-volatile polarization of the ferroelectric.The polarization orientation depends on the polarity of the bias.The remanent polarization either causes accumulation or depletion of the carriers in the channel underneath.⃗ P is the ferroelectric polarization.For applications in programmable photonics, the change in ferroelectric polarization will correspond to a change in the Pockels coefficient of the ferroelectric materials.Hence, if the p-type silicon is etched into a waveguide and a bias is applied across the ferroelectric materials, the effective index of the waveguide mode will be changed in a non-volatile way due to the Pockels effect.instead of material phase transition or free carrier effect leads to an ultra-low loss of 0.07 dB, which is attractive for photonic quantum computing, which requires very low loss.A highly reliable 8-level operation was also demonstrated.Compared to the other technologies, the main drawback of the ferroelectric phase shifter is that only the change in Pockels coefficient is non-volatile, not the refractive index change, which necessitates a constant bias to read the optical state of the phase shifter.A leakage current of ∼200 nA is reported in the recent work, leading to 384 nW of static power in the ON state. 34econd, Pockel's effect is weaker than PCMs or the free carrier effect, leading to a long Lπ of 1 mm.In addition, the low Curie temperature (∼120 ○ C) 98 of BTO may cause erasure of the memory during device thermal cycling.Finally, the hysteretic behavior of ferroelectric materials means that any polarization state will depend on the previous state.Therefore, in order to reliably reach a certain phase shift, the BTO has to be initialized using a long sequence of pulses (∼1 ms in total) to restore all the ferroelectric domains to the same polarization.

FERROMAGNETIC MATERIALS AND MAGNETORESISTIVE MEMORY
Ferromagnetic materials such as CoFeB have also been used in random access memory due to their non-volatile and switchable magnetization under magnetic fields. 99A typical magnetoresistive random access memory (MRAM) cell is made up of a magnetic tunnel junction (MTJ) connected to a transistor 17 [see Fig. 10].A MTJ is a tri-layer structure consisting of a free layer, a tunnel barrier, and a reference layer.The free layer is the switchable ferromagnetic layer, typically made of CoFeB alloys with varying compositions, that stores the information.The tunnel barrier is a thin nonmagnetic dielectric, for example, MgO or AlOx, that provides a means to switch and read the state of the free layer.The reference layer does not switch during memory operations and is used to provide a reference frame for the free layer magnetization.The electrons of a certain spin orientation (i.e., spin up or spin down) can tunnel across the barrier only if there are empty states of the same spin orientation available across the junction.Therefore, when the magnetization directions of the free and reference layers are parallel, the majority spin electrons (spin up) will fill the free majority states across the junction, and the minority spin electrons (spin down) will fill the free minority states [see Figs.10(a) and 10(b)].This will lead to a larger change in tunneling and, hence, lower resistance.On the other hand, when the magnetization orientations are anti-parallel, the majority of spin electrons on one side will try to fill the minority spin states on the other side, resulting in poor band matching and high resistance [see Fig. 10(b)].It becomes clear that by switching the magnetization orientation in the free layer, we can switch the resistance of the MTJ, which is typically realized using two approaches shown in Figs.10(c) and 10(d).A conventional MRAM cell is normally based on the structure shown in Fig. 10(c), where the free layer magnetic switching is induced by the magnetic field generated by the current injected into the bit and word lines.The switching will only occur when the current passes through both the word and bit lines.
The advantage of such an approach is the high switching endurance because the switching is induced solely by a field effect with no physical change in the material structures, as in PCMs and memristors.However, this approach will lead to huge power consumption from the current required to induce the magnetic field, and the fabrication is complicated.Additionally, the magnetic field can also lead to crosstalk, which can accidently switch a nearby cell.To address the limitations, the field has moved toward the spin-transfer-torque MRAM (STT-MRAM) [see Fig. 10(d)].The magnetization of the free layer can be switched by directly passing current through the MJT, and the magnetization orientation is determined by the polarity of the current and the reference layer. 100This method eliminates the need for a word line and the crosstalk issue, as only the cell where the current passes will be switched.Most importantly, as the cell width scales down, the write current becomes significantly smaller compared to the traditional MRAM.

APPLICATIONS OF MAGNETORESISTIVE MEMORY IN PROGRAMMABLE PHOTONICS
To convert the MRAM into a photonic MRAM, one simply needs to add a magneto-optic (MO) material into the structure of an MRAM cell such that the magnetization switching induces an optical phase shift.This was first demonstrated by Murai et al., who used the ferromagnetic material CoFeB as an integrated tunable magnet and Ce:YIG as the MO layer to induce a non-volatile phase shift in an a-Si waveguide 33 (see Fig. 11).Passing current through the Cr/Au wire will induce a magnetic field, which switches the magnetization orientation of the CoFeB thin film underneath depending on the current polarity.The CoFeB in turn induces a non-volatile magnetic field onto the Ce:YIG MO layer underneath the a-Si waveguide, which causes a refractive index change.Figure 11(b) shows that such a non-volatile MO phase shifter is integrated into a microring resonator, which produces a 0.14 nm resonance shift after passing two opposite polarities of current at 500 mA through the coil.A 2 by 2 Mach-Zehnder interferometer switch was also demonstrated with an extinction ratio of 25 dB.Although endurance is not reported in the paper, high endurance is expected because the magneto-switching does not involve any structural change of the materials.Second, the phase shifter can be driven by low voltages (±0.5 V) at the cost of a large device footprint (1.2 mm).However, the high-power consumption (250 mW) for switching is a clear disadvantage compared to other technologies such as memristors and ferroelectrics.Additionally, the MO effect is weak due to the large spacing between the CoFeB and the MO layer, resulting in a long Lπ of 1.2 mm.The reported insertion loss (∼10 dB) of the MZI switch is also prohibitive for large-scale systems.

OPTICAL MEMORY BASED ON VO 2
Similar to PCMs, VO 2 is another class of phase-transition materials that can be reversibly switched between an insulator and metallic phase via electrical, optical, and thermal excitations (Tc ∼ 70 ○ C). 102 Due to the enormous change in the refractive index (Δn, Δk > 2 near 1550 nm), VO 2 is widely used in programmable photonics with applications ranging from electro-absorptive silicon photonic switches 103 to tunable metasurfaces. 104Despite these advantages, the phase transition of VO 2 is volatile-once the excitation is removed, the material will return to the insulator phase.It was only recently that people discovered a way to realize an optical memory in VO 2 35 on a silicon photonic waveguide.The optical memory consists of VO 2 cladded on top of a Si waveguide contacted by electrodes (see Fig. 12).Under electrical bias, a hysteresis loop is observed in the I-V curve due to the insulator-metal-transition (IMT), where the current suddenly increases at threshold voltage.However, once the bias is removed, the VO 2 returns to the high resistance state as the hysteretic behavior is volatile.To write the memory, optical pulses are used to trigger the IMT while the bias is held near the threshold voltage.If the bias is maintained, VO 2 will stay in the metallic state with high absorption.To erase the memory, the bias is turned off, and the VO 2 returns to the insulator phase.The memory can be read either optically through the waveguide transmission or electrically through the resistance.In this way, a light pulse triggered volatile optical memory is realized.Although the switching energy is very low (∼23.5 pJ), substantial power is consumed by holding the memory cell at a high bias voltage (∼13 mW).

COMPARISON BETWEEN DIFFERENT NON-VOLATILE PHOTONIC TECHNOLOGIES
After reviewing these different non-volatile materials, we compare their pros and cons in Table I.We see that the biggest advantage of PCMs is their compact footprint, requiring only tens of micrometers to attain π phase shift, compared to hundreds of micrometers using other technologies under the same bias level or electric field strength.Second, they are a truly "set-and-forget" technology that does not require any external bias or power to maintain their states.This allows the PCM-based devices to be unplugged from the power source in a static state.On the other hand, insertion loss has been the main limitation of PCMs due to the constraint of Kramer-König relations, which stipulates an increase in loss whenever there is an increase in the real part of the refractive index.Even with the discovery of low-loss PCMs, 48,49 it is still challenging to reduce the loss below 0.1 dB/π.The switching time of PCMs is also relatively slow in photonics, which prevents them from high-speed operation.Overall, we think PCMs are well-suited for applications that only require infrequent reconfiguration, and the switches mainly stay in the static state, such as photonic RAM, trimming photonic resonators, optical routing in PICs, and inference for optical neural networks.
Both flash memory and ferroelectric memory are based on the field effect for reconfiguration, which explains why they have very similar switching speed, energy, and modulation strength.The key difference is that flash memory can be truly "set-and-forget," while ferroelectric memory will require a constant read bias to trigger the EO effect.In addition, flash memory is also more CMOS compatible than ferroelectric materials, which are not supported in most commercial foundries.However, ferroelectric materials induce very low insertion loss compared to the high insertion loss caused by the free carrier absorption in photonic flash memory.Flash memory also tends to require a relatively high driving voltage (>10V) to force the tunneling of charges, while ferroelectric materials can typically be switched at <5V. 93 Another drawback of flash memory is its sensitivity to high energy radiation, 86 which can limit its applications in extreme environments such as aerospace.Finally, flash memory tends to have lower switching speeds in the range of a few hundred microseconds. 21We think flash memory will be well-suited for applications where loss and speed are not critical and the information must be stored in a non-volatile way, for example, photonic RAM, but exposure to UV radiation should be avoided.In contrast, ferroelectric memory will have advantages in quantum PICs where loss is critical, 2 thanks to its wide bandgap and the ability to operate at cryogenic temperatures. 109One exception is ferromagnetic memory, which is based on the field effect 33 yet consumes large amounts of power (∼250 mW).A non-volatile MO phase shifter based on ferromagnetic materials will be ideal for applications that require very high endurance and truly "set-and-forget" operation at low voltages (±0.5V), but the large power consumption and insertion loss will be crucial problems that need to be solved before it can be scaled up.
In terms of memristors, they can provide the highest speed and the lowest switching energy because only nanoscale filaments are switched.Similar to flash memory, they are highly CMOS compatible, as the active layer can simply be dielectrics such as SiO 2 , Al 2 O 3 , or HfO 2 .However, depending on their configuration (plasmonics or photonics), they will have very different performance metrics.Plasmonic memristors can be "set-and-forget" but are very lossy, which limits their application to a single standalone device, such as photon sources 32 or detectors. 82The high loss will also be prohibitive for phase shifters.On the other hand, photonic memristors can be a scalable technology thanks to their lower loss and CMOS compatibility, but they require a constant bias to read the optical state, similar to ferroelectric materials.Such DC power consumption is much smaller than VO 2 but still non-negligible (∼nW to μW), especially when the memristor is in the low resistance state.We think they can be potentially used for photonic in-memory computing, where the weight of the neuron can be updated at high speeds during training.Another application is a spatial light modulator where each pixel in a 2D array can be tuned by a memristor at high speed, while the 2D phase mask can be maintained by applying a constant bias to all the pixels.Finally, although VO 2 optical memory exhibits large switching contrasts and can be tuned at relatively high speed REVIEW pubs.aip.org/aip/apm

TABLE I.
For PCMs, the first number quoted is for crystallization, and the second number is for amorphization.For memristors and flash memory, the first number quoted is for SET, and the second number is for RESET.The Lπ is not quoted for GST, GSST, plasmonic memristors, and VO 2 because the high insertion loss prohibits phase-only modulation.
✓ means yes, × means no, and ⚪ means questionable.In the column "Non-volatile?," ⚪ means not "set-and-forget," i.e., a bias is required to read the optical memory state."Amp." is short for amplitude."Sim." is short for simulation.with low switching energy, the absence of non-volatility is a major limitation as ∼mW of power is consumed to maintain the state.The large insertion loss and high-power consumption make VO 2 a less competitive candidate for non-volatile programmable photonics compared to other technologies.

OUTLOOK
In this section, we discuss challenges that need to be addressed to advance the field and identify future research directions.Different materials will be discussed individually in separate sub-sections.

Phase-change materials
Phase-change materials can play an important role in applications that only require infrequent switching, such as photonic RAM, trimming photonic resonators, optical routing in PICs, and inference for optical neural networks.To further scale up the technology, two crucial bottlenecks must be overcome.First, current photonic devices based on PCMs still suffer from relatively high insertion loss (>1 dB).The loss must be reduced below 0.01 dB per device to be useful in large-scale PICs, which will require either careful device engineering 110 and/or very low-loss PCMs with bandgap energy in the ultraviolet range. 111Second, no commercial silicon photonics foundry has supported PCMs so far.Foundries must be willing to incorporate PCMs into their processes so that they can be produced on a wafer scale.This will require first showing attractive functionalities or improved performance from large-scale integrated systems in industrial or academic research settings, which can draw substantial and long-term investment interest.A short-term solution is to deposit PCMs in a back-end-of-the-line process on foundry fabricated wafers, 106 but opening windows in the oxide cladding near active devices for PCM deposition can be a challenging task at the wafer scale. 112

Flash memory
Flash memory is a technology that can be easily scaled up with the current CMOS process.Particularly, photonic flash memory may be the best approach to realize non-volatile photonic RAM, which harnesses the mature flash memory technology developed in the electronic industry.However, current photonic flash memory still relies on carrier injection for modulation, so it will not be ideal for low-loss components such as phase shifters, whereas the driving voltage remains high (20V) and the speed is slow (a few hundred microseconds).While the high insertion loss can be difficult to circumvent due to the nature of carrier dispersion, many studies have proposed different materials and geometry for photonic flash memory that can solve the high driving voltage issue, for example, by guiding the light directly in the charge trapping layers 92 or using graphene as the floating gate. 87We expect to see experimental demonstrations of these ideas as more researchers start to work in this field.

Memristors
Memristor is a promising non-volatile technology that can enable fast and energy-efficient switching and, hence, have potential applications in in-memory optical computing and high-speed spatial light modulation.Similar to flash memory, it is also entirely based on CMOS compatible materials.We believe photonic memristors will be a more promising solution for scalability compared to plasmonic memristors, but a few limitations must be addressed before they can be a competitive technology.More specifically, the DC power required to read the optical state (∼nW to μW)-particularly in the low resistance state-is much higher than that of ferroelectrics and other technologies that do not require a read bias.To mitigate this, the device can be read using short voltage pulses.In this way, the device does not require a constant voltage bias, and read/write voltage is only applied during read and write cycles.In the case of inference for neural networks, for example, it only needs to be written once and then read whenever the vector-matrix-multiplication is being performed.Second, the insertion loss is still relatively high as it is based on carrier effects.One potential solution is to connect the memristor in series with a phase shifter based on EO materials such as BTO or lithium niobate.A constant bias is applied to the EO phase shifter, while the memristor will act as a non-volatile switch to turn the EO effect on and off.This is possible because the resistance of the memristor will determine how much voltage is dropped on the EO phase shifter and on the memristor.The resistance contrast between the LRS and HRS can be larger than 1000 times.Meanwhile, applying a constant bias to the EO phase shifter will consume minimal power due to the low current leakage.In this way, the advantages of both ferroelectrics and memristors can be combined to enable a fast and non-volatile photonic phase shifter with low loss and near zero static power consumption.

Ferroelectric materials
Two key advantages of ferroelectric materials over other technologies are their very low loss and switching energy.The low loss comes from the fact that the reconfiguration mechanism is based on the EO effect rather than the carrier dispersion effect, while the low switching energy can be attributed to field-induced ferroelectric domain switching instead of current-driven switching.On the other hand, two main challenges need to be overcome before wider adoption of this technology.First, the current BTO ferroelectric materials have poor CMOS compatibility, which has to be grown epitaxially and then transferred to a substrate through wafer bonding. 97Although companies like Intel have already successfully introduced heterogeneous integration into their silicon photonics process, a larger incentive, for example, the potential of using BTO as a Pockels material, is still needed to convince the foundries that it is worth the investment to introduce BTO into the fab line.A recent study has shown that the ferroelectric material Hafnium zirconium oxide (HZO), which is a CMOS compatible material and can be easily deposited via sputtering, exhibits a non-volatile phase shifting effect. 113However, the reversible phase transition and the material's Pockels effect were not reported, while the switching voltage is prohibitively high (210 V).Second, the ferroelectric phase shifter still requires a bias to read the optical state, which can complicate the CMOS control as the switching and read pulses need to be synchronized and temporally separated.A potential future research direction will be to investigate whether ferroelectric domain switching can lead to a refractive index change by itself without requiring read bias.

Ferromagnetic materials
One unique advantage of ferromagnetic materials is their potentially high endurance, as magnetic switching does not involve any material's structural changes.Unfortunately, no literature has shown this ultra-high endurance in experiments.Experimental verification of this high endurance in ferromagnetic materials can have a high impact on the non-volatile photonics research field.The main drawback of ferromagnetic phase shifters is the high switching power (250 mW) and weak optical phase shift due to (1) the large spacing between the CoFeB and the Ce:YIG where the waveguide lies and (2) the small mode overlap between the silicon waveguide and the Ce:YIG.To increase the mode overlap with the MO material, we can leverage the mature Si 3 N 4 platforms, which have a lower index contrast compared to Si and have been shown to have a higher MO figure of merit. 114Meanwhile, instead of placing Ce:YIG beneath the waveguide, one can bond Ce:YIG directly on top of the waveguide. 115By doing this, the CoFeB can then be brought into contact with the MO materials, as the Ce:YIG itself can act as a spacer between the absorptive CoFeB and the optical mode, and the CoFeB can exert a much stronger magnetic field onto the Ce:YIG.
In summary, we have reviewed the recent development of nonvolatile materials for programmable photonics applications.These include phase-change materials, flash memory, ferroelectric materials, ferromagnetic materials, memristors, and VO 2 .We briefly explain the working principles behind these materials and discuss their applications in non-volatile photonics.We compare the advantages and disadvantages of different technologies and identify their target applications.Finally, we discuss the crucial challenges and point out potential future research directions to advance the field.
[Fig.3(b-i)].The use of microrings was later improved by Zheng et al., who achieved 33 dB extinct ratio contrast via critical coupling42 Fig. 3(b-ii)] and quasi-continuous operation depending on the number of applied laser pulses.Rios et al. first showed that optical excitation can be REVIEW pubs.aip.org/aip/apmcoupled into a waveguide for switching PCMs and demonstrated a multilevel photonic memory. 23Wu et al. showed that by patterning the PCMs on subwavelength scales and using atomic-layerdeposited Al 2 O 3 as encapsulation, the switching contrast can be significantly improved by almost threefold 55 [Fig.3(b-iii)].Apart from GST, it has been shown recently that low-loss PCMs Sb 2 Se 3 can also be controlled by a laser in a Mach-Zehnder interferometer where 2π phase-only modulation has been achieved 56 [Fig.3(b-iv)].Although optical switching can reduce the difficulty of fabrication, a truly scalable photonic platform requires electrical switching, 44,57 which harnesses the mature CMOS technology that enables the integration of billions of transistors on a single chip.Kato et al. first used the ITO microheater to electrically control GST on the Si waveguide 58 [Fig.3(a-i)], while Zhang et al. doped the Si waveguide to create microheaters 59 [Fig.3(a-iii)].The doped Si microheater was later modified into a PIN diode microheater that significantly reduces the loss from doping 51 [Fig.3(a-iii)].More recently, graphene was also shown to be a reliable and energy efficient microheater for the electrical control of PCMs on various materials and substrates, including Si 60 and SiO 2 . 61[Fig.3(a-iv)].
FIG. 5.Applications of memristors in programmable photonics.(a) A Au/SiO 2 /ITO plasmonic memristor.77The plots show a clear hysteresis in optical transmission upon switching the plasmonic memristor on and off by sweeping the voltage.Extinction ratios of 12 dB are demonstrated in 10 μm devices at operating voltages of ±2V.(b) Single-atom switching in a Ag/aSi/Pt plasmonic memristor.78Switching of the plasmonic resonance in the 1550 nm wavelength range is shown, leading to a large extinction ratio of 10 dB.(c) Silicon photonic memory based on an GaAs/Al 2 O 3 /Si memristor integrated with a microring resonator.30The TEM shows the heterogeneous integration of III-V materials with Si and a 20 nm bonding oxide layer in between.(d) Silicon photonic phase shifter based on an n-InP/HfO 2 /p-Si memristor.31The I-V curve shows typical memristor switching in the electrical domain, while the transmission spectrum shows the switching of microring resonance in the optical domain.(a) is reprinted with permission from Hoessbacher et al., Optica 1, 198 (2014).Copyright 2014 Optical Society of America.(b) is reprinted with permission from Emboras et al., Nano Lett.16, 709 (2016).Copyright 2016 American Chemical Society.(c) is reprinted with permission from Tossoun et al., ■ ■, ■ (2023).Copyright 2023 Author(s), licensed under a Creative Commons Attribution 4.0 License.(d) is reprinted with permission from Fang et al., ■ ■, ■ (2023).Copyright 2023 Author(s), licensed under a Creative Commons Attribution 4.0 License.
FIG. 7.Applications of flash memory in programmable photonics.(a) A silicon photonic read-only memory based on a Si 3 N 4 floating gate.20A microring-based photonic memory was first proposed in simulation, which shows that optical information can be stored in different microring resonance states, controlled by the number of trapped charges.(b) Experimental demonstration of a multilevel programmable silicon photonic memory based on a poly-Si floating gate.21The resonance wavelengths of the on and off states were shown to be stable over 20 hours, and the device can be reversibly switched for 30 cycles.(c) An optically written silicon photonic flash-memory.86UV light is used to optically excite the free carriers into the Si 3 N 4 floating gates.(d) A silicon photonic flash memory with a graphene floating gate.87Graphene can potentially give a larger effective index change per unit of trapped charges compared to poly-Si.(e) A silicon photonic memory with an ITO floating gate.88ITO can be doped by the trapped carriers into the epsilon-near-zero regime, where a large extinction ratio over 10 dB can be realized for a 5 μm long electro-absorption switch.(f) A III-V-on-Si photonic flash memory based on a Mach-Zehnder interferometer.89The dielectric layer consists of alternating layers of Al 2 O 3 and HfO 2 , as revealed by the TEM.The multilayer structures were shown to exhibit more reliable switching compared to pure Al 2 O 3 .(a) is reprinted with permission from Barrios and Lipson, J. Lightwave Technol.24, 2898 (2006).Copyright 2016 IEEE.(b) is reprinted with permission from Song et al., Sci.Rep. 6, 22616 (2016).Copyright 2016 Author(s), licensed under a Creative Commons Attribution 4.0 License.(c) is reprinted with permission from Grajower et al., Laser Photonics Rev. 12, 1700190 (2018).Copyright 2018 Wiley-VCH.(d) is reprinted with permission from Li et al., IEEE Photonics Technol.Lett.28, 284 (2016).Copyright 2016 IEEE.(e) is reprinted with permission from Parra et al., Opt.Lett.44, 3932 (2019).Copyright 2019 Optical Society of America.(f) is reprinted with permission from Cheung et al., ■ ■, ■ (2023).Copyright 2023 Author(s), licensed under a Creative Commons Attribution 4.0 License.

FIG. 9 .
FIG. 9. A non-volatile ferroelectric phase shifter based on BTO.By switching the ferroelectric domain with a sequence of voltage pulses, the Pockels coefficient of the BTO can be changed in a non-volatile way.Under a constant 2V read bias, different phase shifts can be induced due to the Pockels effect, and hence, the microring resonance can be shifted depending on the number of applied pulses.The figure is reprinted with permission from Geler-Kremer et al., Nat.Photonics 16, 491 (2022).Copyright 2022 Springer Nature.

FIG. 10 .
FIG. 10.Operation principle of magnetoresistive random access memory.(a) A magnetic tunnel junction (MTJ) in either parallel or anti-parallel magnetization states.(b) Band diagrams of the MTJ in parallel and anti-parallel states.(c) Schematic of a conventional MRAM cell.The magnetic pillar corresponds to the MTJ shown in (a).(d) Schematic of a STT-MRAM.Figures (a) to (d) are reprinted with permission from A. Makarov, ■ dissertation (■, ■).101

FIG. 11 .
FIG. 11.A non-volatile magneto-optic phase shifter.(a) Schematic showing the structure of the magneto-optic phase shifter.The magnetization of the CoFeB can be switched by passing current through Cr/Au electrodes.Ce-YIG is a magneto-optic material whose refractive index changes in response to the magnetic field induced by CoFeB.(b) A microring integrated with the magneto-optic phase shifter.A non-volatile magneto-optic wavelength shift of 0.14 nm can be induced by passing 500 mA current through the electrodes.The figures are reprinted with permission from Murai et al., Opt.Express 28, 31675 (2020).Copyright 2020 Optical Society of America.

FIG. 12 .
FIG. 12. Optical memory based on VO 2 .When the memory cell is held at a constant bias of 9 V in the high transmission and resistance state, it can be switched into the low transmission and resistance state by a 100 ns optical pulse.The cell can be returned to its high transmission and resistance state by removing the voltage bias.The optical contrast can be increased by using a longer VO 2 segment.The figure is reprinted with permission from Jung et al., ACS Photonics 9, 217 (2022).Copyright 2022 Author(s), licensed under a Creative Commons Attribution 4.0 License.