Light storage in light cages: a scalable platform for multiplexed quantum memories

TL;DR

Researchers developed a scalable quantum memory platform using 3D-printed light cages to store light pulses efficiently in cesium vapor.

Contribution

The work introduces 3D-nanoprinted hollow-core waveguides (light cages) for scalable, chip-integrated quantum memories with improved performance.

Findings

Light cages enabled hundreds of nanoseconds of photon storage in cesium vapor.

Multiple LC memories were integrated on a single chip with consistent performance.

The platform shows potential for parallel single-photon synchronization in quantum computing.

Abstract

Quantum memories are essential for photonic quantum technologies, enabling long-distance quantum communication and serving as delay units in quantum computing. Hot atomic vapors using electromagnetically induced transparency provide a simple platform with second-long photon storage capabilities. Light-guiding structures enhance performance, but current hollow-core fiber waveguides face significant limitations in filling time, physical size, fabrication versatility, and large-scale integration potential. In this work, we demonstrate the storage of attenuated coherent light pulses in a cesium (Cs) quantum memory based on a 3D-nanoprinted hollow-core waveguide, known as a light cage (LC), with several hundred nanoseconds of storage times. Leveraging the versatile fabrication process, we successfully integrated multiple LC memories onto a single chip within a Cs vapor cell, achieving consistent performance across all devices. We conducted a detailed investigation into storage efficiency, analyzing memory lifetime and bandwidth. These results represent a significant advancement toward spatially multiplexed quantum memories and have the potential to elevate memory integration to unprecedented levels. We anticipate applications in parallel single-photon synchronization for quantum repeater nodes and photonic quantum computing platforms.

We implement 3D-nanoprinted hollow-core waveguides—so-called light cages—as atomic vapor-based quantum memories. These structures significantly enhance light–matter interactions within a compact, chip-integrated platform, marking a step forward in scalable and versatile photonic quantum technologies.

Figures

Spatially multiplexed quantum memory using hollow-core light-cages (LC) on a chip.a Illustration of several LCs guiding light through their core immersed in a Cs atmosphere. b Scheme of the use of LCs as a spatially multiplexed on-chip quantum memory (see text). c SEM image of four LCs printed on the same chip, with two different geometries, single-layer and dual-layer of strands. Printing several waveguides hundreds of micrometers apart in the same printing process is possible. d Zoom-in of a 12-strand LC. Top inset: Cross section of the LC with the relevant design parameters, *D*_*s*_ = 3.6

Experimental setup.a EIT is created using co-propagating signal and control fields, coupled into the LC with a pair of aspheric lenses (f). The control field is filtered using polarization and frequency filtering. Pulses are carved from CW lasers using a pair of EOMs. (OD: neutral density attenuators, Λ*/*2 half-wave plate, Λ*/*4 quarter-wave plate, PBS polarizing beam splitter, FPx tunable Fabry-Pérot cavities, EOM electro-optic modulator, AWG arbitrary waveform generator, SNSPD superconducting nanowire single-photon detector, TCSPC time-correlated single photon counting unit). b Energy level

EIT in the LC.a Measured (solid line) EIT transmission spectra of light coupled into the LC at different control powers. The EIT peaks are visible on the two-photon resonance (∆*ν*_*s*_ = 0) with increasing widths and transmission. Theoretical fits are plotted as dashed lines. b The Rabi frequency and EIT contrast (*OD*_EIT_) as obtained from the fits at varying control powers (see details in the main text)

Storage of light in the LC.a The leak pulse (filled in orange), i.e., the part of the input pulse that escaped storage, sets the time zero. Arrows indicate the pulse width. The retrieved read pulse (filled in green) appears after 52 ns. Experimental parameters are: *T* = 74.0 °C, input signal pulse width ∆*t*_*s*_ = 14.1(4) ns, coupling power *P*_*c*_ = 10.0(1) mW. The internal efficiency for 52 ns storage is *η*_int_ = 0.097(1). b Memory lifetime measured for different storage times. A 1*/e*-lifetime of 83(2) ns is obtained using a heuristic fit that considers the precession of electronic spi

Memory efficiency as a function of lifetime and bandwidth.a Heat map of the intensity of the read signal for varying set storage times with *P*_*c*_ = 10.0(1) mW, ∆*t*_*s*_ = 14.1(4) ns. An oscillatory behavior due to spin precession is observed. b Storage efficiency as a function of storage time for different control powers. The internal efficiency reduces with increasing powers for shorter storage times. c Heat map of the intensity of the read signal with varying input signal pulse widths, ∆*t*_*s*_, with a fixed set storage time of 150 ns. d Storage efficiency as a function of the signal pu

Storage of light in two different LCs on a chip.Storage efficiency measurements were performed in two neighboring LCs with the same geometrical parameters (LC-A and LC-B) on a single chip. Control powers and signal pulse widths were *P*_*c*_ = 20.0(3) mW and ∆*t*_*s*_ = 14.2(4) s, respectively. Both LCs perform nearly identically within the fitting errors, *t*_mem__−__*A*_ = 86(3) ns and *t*_mem__−__*B*_ = 87(3) ns. Inset: Schematic of the coupling geometry

Full text

Introduction

The ability to store quantum information is crucial for various quantum technologies. In long-distance quantum communication, signal loss in quantum channels poses a fundamental limitation^1^. Quantum repeaters^2^ mitigate this challenge through entanglement swapping, where quantum memories play a key role by enabling teleportation and synchronizing entanglement-swapping operations^3,4^. Experimental studies demonstrate that incorporating quantum memories in repeater segments significantly enhances photon pair rates compared to memory-less implementations, using quantum memories with storage times ranging from hundreds of nanoseconds to tens of milliseconds^5,6^. Beyond communication, quantum memories are essential in photonic quantum computing, providing photon synchronization and controllable delays necessary for feed-forward operations in measurement-based quantum computing^7,8^. Experimental demonstrations have shown the creation of linear cluster states with fidelities *>*90% using memories with storage times in the microsecond scale^9^.

Quantum memories have been developed across various platforms, including solid-state systems^10–13^, ultra-cold atoms^14–16^, and hot atomic vapors using electromagnetically induced transparency (EIT)^17–21^. Hot atomic vapors offer low technical complexity, making them attractive for scalable quantum memories. Long storage times of hundreds of milliseconds have been demonstrated in cesium vapor using coherent light pulses^17^. It is noteworthy that these long storage implementations have thus far been performed with classical light.

Combining atomic vapor quantum memories with waveguiding structures enhances light-matter interaction, reducing the memory’s footprint and the optical power required to operate it. Recent experimental implementations have demonstrated the storage of light via EIT using atoms within the core of optical waveguides, predominantly employing hollow-core fibers (HCFs)^22–27^. However, these approaches face significant challenges. Loading atomic vapors into the cores of HCFs is typically done through the fiber’s end facets, resulting in filling times that can span months^22,27^. Alternatively, laser-cooled atoms have been used to load the fibers^23–26^, but this introduces additional experimental complexity due to the need for atom trapping and cooling. Moreover, the fiber-pulling fabrication process constrains the shape and design of hollow-core fiber memories. HCFs are not easily integrated into optical chips, further limiting their practicality.

To address these limitations, we focus on the unique versatility of so-called hollow-core light cages (LCs) to realize quantum memories with a small footprint, great flexibility in design, and easy integration into photonic chips. LCs are anti-resonant hollow core waveguides with unique side-wise access to the core region^28^. These on-silicon-chip structures, 3D-nanoprinted from commercially available photoresins using two-photon polymerization, have shown their worth in optofluidic analysis^29^, spectroscopy^30^, and nanoparticle characterization^31^. Additionally, the LCs show a notable enhancement in the diffusion of molecular gases^28^, as well as atomic gases^32^, compared to the traditional HCFs, reducing the time needed to reach reasonable optical depths, from months to days, while keeping the optical fields confined in the LC core. By tuning the geometrical parameters of the LCs, the anti-resonances can be tuned to guide light across specific spectral intervals in the visible spectrum and the near-infrared^28,30,33^. Coherent light-matter interaction has been shown by creating EIT inside LCs placed in a Cs vapor cell and achieving Rabi frequencies in the order of hundreds of MHz^32^.

In this work, we demonstrate the use of LCs as on-chip quantum memories based on EIT, achieving storage times of several hundred nanoseconds. We integrated multiple LC memories onto a single chip within a cesium (Cs) vapor cell and observed consistent performance across all devices. This marks a significant step toward spatially multiplexed quantum memories and could scale up memory integration to a new level. We anticipate applications in parallel single-photon synchronization for quantum repeater nodes and photonic quantum computers. Our thorough investigation includes the storage efficiency of faint coherent light pulses in LCs and an analysis of the memory’s lifetime and bandwidth, emphasizing their dependence on control power during writing and reading processes. Moreover, we examine the current limitations of LC-based quantum memories, especially concerning efficiency and decoherence times.

Results

Light cages in atomic vapor

A LC is a special type of anti-resonant hollow-core waveguide^34^ that guides light along its core while allowing for fast diffusion of Cs inside the core through its unique side-wise access^32^, as illustrated in Fig. 1a. The conceptual idea of a spatially multiplexed on-chip quantum memory is schematically shown in Fig. 1b. Flying qubits encoded in the polarization or time-bins of individual photons from single photon sources (SPS) are separated by beam splitters (polarization beam splitters or switches) and fiber-coupled to several LCs on a chip inside an atomic vapor cell. After storage and synchronization by choosing different storage times for each LC the photons are released, coupled into fibers, and partially processed, e.g., by a Bell state measurement (BSM). Such a functionality is needed, e.g., in quantum repeaters.Fig. 1. Spatially multiplexed quantum memory using hollow-core light-cages (LC) on a chip.a Illustration of several LCs guiding light through their core immersed in a Cs atmosphere. b Scheme of the use of LCs as a spatially multiplexed on-chip quantum memory (see text). c SEM image of four LCs printed on the same chip, with two different geometries, single-layer and dual-layer of strands. Printing several waveguides hundreds of micrometers apart in the same printing process is possible. d Zoom-in of a 12-strand LC. Top inset: Cross section of the LC with the relevant design parameters, Ds = 3.6 µm, Λ = 7 µm, D = 28 µm. Bottom inset: Collimated output beam. x-y beam profiles fitted to Gaussian functions (yellow dashed lines) showing that the fundamental mode TEM_00_ is coupled in the LC

The LCs were designed and 3D-nanoprinted on a silicon substrate following the two-photon polymerization technique described in ref.^28^. (see materials and methods). A detailed overview of key properties of the LC structure investigated in this work, compared to benchmark values reported in the literature^35^, can be found in the supplementary information S1. A scanning electron microscope (SEM) image in Fig. 1c shows several fabricated LCs on a chip. A zoom-in of an individual LC and the measured mode profile of guided light is depicted in Fig. 1d. The LC chip is placed inside a Cs vapor cell, and light is coupled into individual LCs. To prevent degradation of the polymer LCs by the chemically reactive Cs gas, the structure is coated with 100 nm of alumina. We found remarkable stability and observed no indication of degradation even after 5 years in the Cs atmosphere. The coating also serves to fine-tune the waveguide resonances^36^ to the Cs D_1_ line at 894 nm.

The experimental setup of the quantum memory is shown in Fig. 2a. Signal and control laser beams are generated from independent diode lasers locked to the transitions F = 3 − F' = 3 and F = 4 − F' = 3 of the Cs D_1_ line, respectively (see energy level diagram in Fig. 2b). The beams are coupled collinearly with orthogonal polarizations into the LCs on the chip inside the vapor cell using a pair of aspherical lenses. We found a transmission efficiency ηLC = 0.20(1). The in-coupling efficiency ηcoup, as defined in ref.^29^ is calculated using the measured modal attenuation α = 1.19(3) dB*/mm before placing the chip inside the vapor cell, resulting in ηcoup = 0.82(4) through the vapor cell (see supplementary information S2 for details on the calculation). After transmission, the control beam is suppressed through polarization filtering, and the signal is detected by a superconducting nanowire single-photon detector (SNSPD). A key part of the memory, the vapor cell, is heated to T = 74 °C to increase its optical depth, and the transmission spectra are measured while scanning the signal laser frequency (see materials and methods for more details).Fig. 2. Experimental setup.a EIT is created using co-propagating signal and control fields, coupled into the LC with a pair of aspheric lenses (f). The control field is filtered using polarization and frequency filtering. Pulses are carved from CW lasers using a pair of EOMs. (OD: neutral density attenuators, Λ/2 half-wave plate, Λ/4 quarter-wave plate, PBS polarizing beam splitter, FPx tunable Fabry-Pérot cavities, EOM electro-optic modulator, AWG arbitrary waveform generator, SNSPD superconducting nanowire single-photon detector, TCSPC time-correlated single photon counting unit). b Energy level of the employed Λ-system in the Cs D_1_ line. Signal (ω_s) and control (ωc) fields are locked to the transitions F = 3-F' = 3 and F = 4-F' = 3. The decoherence rate between the ground states F = 3 and F = 4 is denoted as γd*_

In Fig. 3a, we can observe the EIT transmission peak at the two-photon resonance (signal detuning ∆νs = 0) for different powers of the control beam. With increasing power, the transmission within the transparency window increases and approaches nearly unity at Pc = 40 mW. The apparition of “shoulder-like” features at ∆νs ∼ −0.8 GHz and 1.6 GHz is caused by the propagation loss in the LC. These features confirm that the observed EIT stems from light propagation in a mode of the LC. The spectra are fitted by the model proposed in ref.^32^, using a propagation-dependent linear susceptibility, χ(z), arising from the attenuation of the control Rabi frequency as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Omega }_{c}\left(z\right)={\Omega }_{c}\left(0\right){e}^{{\rm{\alpha }}z}$$\end{document} , with α being the modal attenuation of the LC (see supplementary information S3 for details on the propagation dependent model of the transmission). Using this model, we can extract Ω_c, which is plotted in Fig. 3b, as a function of the control power. Here we observe an increasing Ωc_ for rising control powers, showing a maximum value of Ω_c_ = 232(43) MHz. The corresponding transparency window width ∆fEIT = 133(24) MHz is computed from Ω_c_ after considering the Doppler broadening and optical depth of the atomic medium. On the other hand, the EIT contrast (ODEIT), defined as the reduction in optical depth at the EIT resonance^37^, remains almost constant with a value of ODEIT ≈ 1. This shows the LC is capable of handling high control laser powers producing broad transparency windows, a requirement for storage of broadband light pulses.Fig. 3EIT in the LC.a Measured (solid line) EIT transmission spectra of light coupled into the LC at different control powers. The EIT peaks are visible on the two-photon resonance (∆νs = 0) with increasing widths and transmission. Theoretical fits are plotted as dashed lines. b The Rabi frequency and EIT contrast (ODEIT) as obtained from the fits at varying control powers (see details in the main text)

Storage of light in a light cage

To implement the EIT storage protocol, we cut pulses out of the continuous wave signal and control beams with a pair of amplitude electro-optic modulators (EOM). A first control pulse performs optical pumping. Then, a faint signal pulse with 2/3 the length of the control pulse is sent simultaneously with a second control pulse. This transfers the signal pulse into a spin wave in the atomic ensemble via EIT. After a set storage time tstorage, the process is reversed, i.e., the stored light is released by sending a third control pulse.

Figure 4a shows a measurement using a peak control power of Pc = 10 mW and an attenuated signal pulse characterized by an average photon number of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\left|\alpha \right|}_{{\rm{in}}}^{2}=50(5)$$\end{document} , and a pulse width (FWHM) of ∆ts = 14.1(4) ns. The average control noise photons per pulse is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\left|\alpha \right|}_{{\rm{noise}}}^{2}=0.02(1)$$\end{document} , measured from the control field without signal field present, resulting in a high signal-to-noise ratio, SNR > 2000 (details on the photon average estimations can be found in the supplementary information S4). The first prominent peak from light that escaped storage (“leak”) defines the time zero in the plot. We define the internal memory efficiency as the ratio of the integrated counts from the retrieved signal (“read”) to the total amount of counts, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\eta }_{\mathrm{int}}={N}_{{\rm{read}}}/\left({N}_{{\rm{read}}}+{N}_{{\rm{leak}}}\right)$$\end{document} . A temporal filter of width 3σin is applied to both pulses, where σin is the variance of the fitted Gaussian distribution to the leak pulse (shaded areas in Fig. 4a). For a storage time of 52 ns, the internal memory efficiency is ηint = 0.097(1). This shows that the LC can store light as a spin-wave excitation in the atomic medium even in a regime where ODEIT is relatively low.Fig. 4. Storage of light in the LC.a The leak pulse (filled in orange), i.e., the part of the input pulse that escaped storage, sets the time zero. Arrows indicate the pulse width. The retrieved read pulse (filled in green) appears after 52 ns. Experimental parameters are: T = 74.0 °C, input signal pulse width ∆ts = 14.1(4) ns, coupling power Pc = 10.0(1) mW. The internal efficiency for 52 ns storage is ηint = 0.097(1). b Memory lifetime measured for different storage times. A 1*/e*-lifetime of 83(2) ns is obtained using a heuristic fit that considers the precession of electronic spins produced by stray magnetic fields (see text). The shaded blue area indicates the 95% confidence interval of the fit. Inset: Selected set of measurements analyzed to calculate the memory lifetime. An intensity offset is added for better visibility

We then studied the dependency of the memory efficiency ηint on the set storage time. The results are plotted in Fig. 4b. Surprisingly, the efficiency does not show the expected mono-exponential decay but a damped oscillatory behavior. We attributed this to a spin precession due to residual magnetic fields while the light is stored as a spin-wave. We fit the data using an exponentially damped sinusoidal oscillation, where the oscillation frequency is determined by a magnetic field. The fit suggests a magnetic field that agrees well with the field expected from the Earth and heating wires (see supplementary information S5). The observed spin precession enables a novel way of magnetic polarization control of the stored light pulses, an application that will be studied further in future works.

As a next step, we investigated the influence of the control power Pc on memory efficiency. The heat map in Fig. 5a shows the intensity of the read signal (retrieved pulse after storage) as a function of the measured and actual set storage times for a signal pulse width of ∆ts = 14.1(4) ns and Pc = 10.0(1) mW. Again, the oscillatory behavior due to spin precession is observed. From the diagonal cross sections of such heat maps for different control powers, the storage efficiency ηint as a function of the storage time is derived and plotted in Fig. 5b. Counterintuitively, ηint decreases with increasing Pc for shorter storage times. This can be attributed to the power-dependent spatial compression of the input signal pulse while traveling through the LC in the atomic vapor cell. With increasing control power, the EIT transparency window broadens, but the group velocity decreases. Following refs. ^32,38^, we simulated the group velocity of a pulse propagating through the cell (see supplementary information S3). At Pc = 10 mW, a pulse compression by a factor of 143 occurs, i.e., a 14 ns signal pulse traveling through the LC is compressed to a length of 29 mm. As the optical length of the LC is 5 mm, only a fraction of the pulse is stored. Increasing Pc reduces the compression further and leads to even longer signal pulses in the LC. For this reason, increasing Pc reduces the storage efficiency in our scenario. This effect becomes less pronounced for longer storage times where the intrinsic loss mechanisms, such as decoherence by atomic collisions, dominate. To obtain a higher efficiency the LC has to be extended or the optical depth has to be increased.Fig. 5. Memory efficiency as a function of lifetime and bandwidth.a Heat map of the intensity of the read signal for varying set storage times with Pc = 10.0(1) mW, ∆ts = 14.1(4) ns. An oscillatory behavior due to spin precession is observed. b Storage efficiency as a function of storage time for different control powers. The internal efficiency reduces with increasing powers for shorter storage times. c Heat map of the intensity of the read signal with varying input signal pulse widths, ∆ts, with a fixed set storage time of 150 ns. d Storage efficiency as a function of the signal pulse width for different control powers. The storage bandwidth increases with increasing control power

Finally, we characterize the bandwidth of the LC-memory by setting a fixed storage time and varying the signal and control pulse widths, where we keep the ratio of the control to the signal width constant, i.e., ∆tc = 1.5∆ts. The results are shown in Fig. 5c, d. The shift between measured time and set storage time arises from the changes in the control pulse width. Broader control pulses result in earlier read pulses (see supplementary information S4). An important parameter of a memory is its bandwidth. It is defined as ∆fBW = 1*/∆t_s_ for a value of ∆t_s_ where the storage efficiency ηmem is −3 dB and can be obtained from the fits. Here, we find a bandwidth of 35.2(6) MHz. These results indicate that the optimal operating point for the LC-memory is found in the low P_c*_ regime. Importantly, we can utilize higher power levels than necessary without compromising the transmission through the LC. This capability paves the way for investigating non-linear processes that require high-intensity light fields guided within the LCs without causing damage.

Spatial multiplexing of light cage memories on a chip

As indicated in Fig. 1, spatial multiplexing requires the fabrication and integration of several identical quantum memories. We demonstrate this conceptually by printing several LCs on the same chip, only hundreds of µm apart, and storing light pulses therein. The nanoprinting process allows precise and reproducible fabrication of complex on-chip hollow-core waveguide structures. Extensive studies show extraordinarily high intra-chip reproducibility with structure variations of less than 2 nm and minimal inter-chip variations of less than 15 nm^35^, both of which are crucial for the multiplexing concept discussed here. As an example, the storage efficiency at a fixed pulse width of 14.1 ns was measured as a function of the storage time for two different LCs, labeled LC-A and LC-B, see Fig. 6. The performance of both LC-A and LC-B is identical, with the same storage times of 86(3) ns and 87(3) ns, respectively. The small difference in the storage efficiency is likely due to a different lens coupling efficiency of the signal and control fields to the LC. This shows that the employed nanoprinting technique creates LCs with replicable optical properties, laying the groundwork for spatial multiplexing using an array of LC-memories on the same chip.Fig. 6. Storage of light in two different LCs on a chip.Storage efficiency measurements were performed in two neighboring LCs with the same geometrical parameters (LC-A and LC-B) on a single chip. Control powers and signal pulse widths were Pc = 20.0(3) mW and ∆ts = 14.2(4) s, respectively. Both LCs perform nearly identically within the fitting errors, tmem_−A_ = 86(3) ns and tmem_−B_ = 87(3) ns. Inset: Schematic of the coupling geometry

Discussion

Nanoprinted hollow-core LCs in an atomic vapor environment are suitable for storing light using EIT for hundreds of nanoseconds. A figure of merit for comparing our LC-memory to other hollow-core light-guiding structures is the fractional delay, F. It is defined as the ratio of storage time to signal pulse width, F = tstorage/∆ts, and can be used to evaluate the performance of the memory. In our system, we achieved a fractional delay close to 4 for pulses with a 14 ns temporal width. This is comparable to results obtained with hollow-core photonic crystal fibers (HCPCFs). Peters et al.^24^ achieved F ≈ 3 using Rubidium atoms from a magneto-optic trap, while Sprague et al.^22^ and Rowland et al.^27^ both reached F ≈ 10 with λ- and ladder-systems, respectively, where the latter is suitable for broader bandwidth. We like to point out that our systems utilize warm atomic vapor, which fills the LCs immediately, while HCFs need years to reach filling saturation. Further, the design of hollow-core fibers is limited by the fiber-pulling process, whereas the printing of LCs by two-photon polymerization allows for arbitrary three-dimensional shapes and rapid integration on a chip. The alumina-coated structures are well-suited to withstand the reactive atmosphere of Cs, showing no degradation of their transmission even after 5 years of being sealed in their vapor cells.

We observed that the LC-memory preserves the spin coherence of the atomic vapor in its interior long enough to observe spin precession in residual magnetic fields. With external magnetic field coils, this can be precisely controlled^37^ and utilized to fine-tune the polarization of the stored light pulse. Such a feature helps to improve filtering, but also is crucial for constructively interfering stored and retrieved light pulses as recently demonstrated in efficiency enhancement via light-matter interference by Burdekin et al.^39^. The observation of spin precession is a direct indication of the coherence of the storage process. Although storage by electromagnetically induced transparency is perfectly described purely classically^40^, this is crucial for any “quantum application” of a memory.

The reproducibility of the storage performance of different LCs with the same geometrical parameters together with the small footprint of warm atomic vapor systems opens the possibility of spatial multiplexing with several LC-quantum memories on the same chip. Exploiting the field confinement inside the LC core, together with nanoprinted fiber-coupling technologies developed for these structures^33^, as well as fiber feedthroughs into vapor cells^41^, could scale up quantum memory integration and technology to a new level.

Frequency multiplexing has been shown in integrated laser-written rare-earth doped waveguides, able to store hundreds of temporal modes^42^, and achieving long storage times up to a millisecond^43^, despite these remarkable properties, the bandwidth of each mode in these memories is ~1 MHz, with the additional complexity of cryogenic systems to cool down the waveguides to 3 K. Here our proposed memory system, operating slightly above room temperature excels in practicality with higher bandwidths per mode. To increase even further the bandwidth of our memory, Raman protocols can be implemented to achieve ~100 MHz^44^. It is essential to note that these Raman protocols require a detuning of the control and signal fields up to 2 GHz to reach an optimal efficiency^45^. Due to the wide anti-resonance window (80 nm) at our target wavelength, implementing Raman protocols is straightforward. Conversely, vapor memories at room temperature struggle to achieve the ultra-long storage times needed, such as for quantum applications repeaters. We envision the applicability of our platform towards controllable delay stages and photon synchronizers. Synchronization of larger non-classical states, such as squeezed states, is required, e.g., to facilitate the generation of heralded non-Gaussian states^46^. Here, noise is also crucial, but less than in experiments where high-quality single photon states need to be stored. The compact size and compatibility with laser-written structures benefit these applications. Thus, a complementarity exists among the various platforms.

Despite the advantages of the presented platform, the LC-memory still needs improvement to surpass macroscopic vapor cells in terms of internal efficiencies and fractional delays. The measured optical depth ODEIT ≈ 1 is below the expected OD compared to the free space scenario. By measuring the optical depth for a focused beam in a 5 mm vapor cell, without LCs, we obtain ODEIT = 3.5(5), indicating the atomic density inside the LCs is effectively reduced, most probably due to the adsorption of Cs atoms on the walls of the waveguide. This could be compensated by increasing the temperature, but the current structure has a measured critical temperature of 80 °C. Beyond this point, the coupling efficiency of the structure degrades rapidly due to the inelastic deformation of the polymer at higher temperatures. Increasing the length of the LCs and employing light-induced atomic desorption techniques (LIAD)^22,47^ would increase the optical depth in the waveguides significantly without the need to increase the temperature.

Several approaches can be implemented to enhance the storage times. Our primary limitations stem from the decoherence of the ground states due to insufficient magnetic field control. Additionally, given the small core diameter, the time-of-flight (TOF) of the atoms inside the LC becomes a significant decoherence process, taking on the order of hundreds of nanoseconds for an atom to traverse the core region transversely (refer to supplementary information S6 for details). Modifying the heaters to minimize induced magnetic fields, along with implementing active field compensation and passive magnetic shielding, would lead to longer coherence times and, consequently, longer memory lifetimes, extending well into the microsecond range^48^. Furthermore, the use of buffer gases, such as nitrogen or neon, would reduce spin-exchange collisions, enhancing storage times to milliseconds^20^. The use of buffer gases along with an increase in the LC core diameter will reduce TOF decoherence, enabling the memory to operate in the millisecond storage regime. Moreover, targeting specific sub-Zeeman levels to harness the Zeeman coherence ∆m = 1 and achieving a spin-exchange relaxation-free condition can increase the memory lifetime to hundreds of milliseconds^17^. A novel approach that utilizes constructive light-matter interference within the memory to enhance the storage efficiency by more than threefold can also be integrated into our system^39^. The required interferometric setup can be simplified on our platform through direct fiber coupling to a memory-on-chip.

So far, our experiments have been conducted utilizing attenuated classical light pulses. However, the enhancement of storage efficiency, reduction of noise, and preservation of the coherence of stored light are synonymous within both the quantum regime and the ultra-low-intensity classical pulse regime. This indicates that a vapor memory, which is proficient in efficiently and coherently storing classical pulses with an average photon number of one, will similarly function effectively for non-classical light, such as single photons.

The transmission properties of the LC are crucial for the multiplexing concept presented here. The optical losses for the fabricated structure are approximately 1 dB/mm^35^, which is acceptable for this study but may need to be reduced if longer structures are considered. Increasing the number of rings of strands surrounding the structure, as shown in Fig. 1b, could lead to an enhancement of −21.7 dB*/ring^49^. Future optimization strategies will focus on minimizing waveguide roughness, which currently limits further loss reduction. These strategies include annealing and advanced writing techniques, such as grayscale lithography^50^. Achieving a 100-fold reduction in the current modal attenuation would result in an enhancement of the control Rabi frequency by a factor of 5, i.e., Ω_c*_ = 1 GHz, while allowing for the LC to be extended. Increasing the length to 20 mm would boost the optical depth up to 4, thereby improving the storage efficiency accordingly without requiring more control power. Additionally, increasing the air-filling fraction through local structuring, as shown in ref.^31^, could further mitigate the impact of polymeric surfaces and enhance performance by minimizing the observed atomic density reduction inside the waveguides, increasing the optical depth to OD > 10, while using a 20 mm LC without the need to increase the temperature of the vapor cell above 70 °C.

Improving the thermal stability of the LC structure represents another important future research direction. Several advanced material systems offer promising solutions. GP-Silica, a recently developed photoresin by Nanoscribe, enables the 3D printing of silica glass microstructures with exceptional thermal stability above 1000 °C after high-temperature sintering^51^. OrmoComp, a hybrid organic-inorganic resin, combines polymer processability with enhanced mechanical and thermal properties and has been used for fabricating complex micro-optical components with improved stability^52^. Additionally, polyimide-based nanoprinting allows the fabrication of sub-micrometer structures that remain mechanically robust at temperatures exceeding 300 °C^53^. These material platforms will be explored in future work to further expand the operational temperature range of nanoprinted LC devices.

Another key aspect of advancing this concept is the photonic integration of structures into on-chip and fiber architectures. Recent advances in the field include fiber-coupled integration of LCs on V-groove functionalized silicon chips for spectroscopy^33^ and direct printing of LCs onto fiber end faces^54^.

In addition to the LC geometry, a recently developed anti-resonant hollow core waveguide features a closed square polymer membrane surrounding the central hollow core gaps every 150 µm for gas access. This structure can be integrated into on-chip environments^55^ or interfaced with optical fibers and offers broad spectral transmission bands with reduced structural complexity, potentially benefiting future multiplexing experiments. Moreover, these waveguides can be interfaced with phase plates that allow selective mode excitation within the waveguide core^56^, further enhancing the flexibility of the concept.

Indeed, a reduction in modal attenuation, an increase in atomic density within the core through longer LCs and higher operating temperatures, as well as a reduction of the surface areas where atomic deposition can occur, would not only improve storage efficiency but also decrease the control power needed for low-noise memory operation, enabling the system to function at the single-photon level. This is complemented by the versatility of a laser-written waveguide with direct fiber coupling and spatial multiplexing capabilities. As a result, our platform could serve as a controllable delay or photon synchronizer, where delays ranging from nanoseconds to several microseconds can notably increase the pair generation rate of heralded single-photon sources^6^.

In conclusion, we have achieved storage of faint light pulses in a LC structure for hundreds of nanoseconds for the first time. The fractional delay, a figure of merit of a light memory, is on par with other memories using light confinement. We demonstrated the versatility and reproducibility of the fabrication process by integrating several LC-memories on a single chip inside a Cs vapor cell. We found nearly identical storage behavior of individual LC. Our results can lead to a plug-and-play on-chip quantum memory or quantum synchronizer with multiplexing capabilities.

Materials and methods

Light Cage: fabrication and integration into an alkali vapor cell

The LC was fabricated by 3D two-photon polymerization lithography on a silicon substrate using a commercial 3D printing system (Photonic Professional GT, Nanoscribe GmbH) and a negative resist (IP-Dip, Nanoscribe GmbH). A 63× objective with NA = 1.4 allowed a sub-micron lateral resolution of 400 nm. A custom General Writing Language (GWL) program was used to optimize implementation conditions to improve mechanical stability. The lateral position of the laser beam was controlled by galvo mirrors at a speed of 55 mm/s. The system operated with a 780 nm NIR femtosecond laser (100 fs pulse width, 80 MHz repetition rate) and an average power of 12.4 mW before entering the objective. The horizontal layer spacing was 150 nm, with 100 nm between lines within a layer. After printing, the structure was coated with alumina using low-temperature ALD at a growth rate of 2.2 ˚A per cycle at 30 °C. The chip was mounted on aluminum and secured with a fused glass clamp. The custom optical-quality glass cell, filled with less than 1 g of cesium under high vacuum, was placed in an oven with resistance heaters and insulated with a polyoxymethylene (POM) housing.

EIT and light storage setup

EIT is created using co-propagating orthogonally polarized signal and control fields, as shown in Fig. 2. Both lasers are locked to their atomic transition through saturated absorption spectroscopy (SAS). The control laser (DL pro and BoosTA, Toptica) undergoes −50 dB suppression through high-quality calcite prisms and additional −60 dB via cascaded Fabry-Pérot etalons with an FWHM of 400 MHz and 800 MHz. The fundamental mode TEM_00_ is effectively coupled into the LC using the aspherical lenses f (A220TM-B, Thorlabs) and inspected with a CMOS camera (Fig. 1d). The LC coupling achieves transmission efficiency ηLC = 0.20(1), with modal attenuation α = 1.51(2) dB*/*mm at 894 nm, including losses due to the modal attenuation, coupling efficiency, and losses on the vapor cell windows. EIT spectra characterization in the LC is performed by scanning the signal laser (EYP-DFB-0894, Eagleyard Photonics) and the fiber-coupled photodiode (DET36A2, Thorlabs) for detection. The storage sequence is executed using a pair of EOMs (AM905b, Jenoptik) with active BIAS stabilization through optical power monitoring. The signal is detected with SNSPDs (Eos CS, Single Quantum) and analyzed with a TCSPC system (PicoHarp 300, PicoQuant).

Supplementary information

Supplementary Information for: Light Storage in Light Cages: A Scalable Platform for Multiplexed Quantum Memories