Nanophotonics-enabled optical data storage in the age of machine learning

Over the last few decades, the digital transformation that our society has undergone has resulted in an unprecedented amount of data being produced. According to the International Data Corporation, global data generation will reach 175 zettabytes by 2025 (1 zettabyte equals 1 × 10⁹ terabytes).¹ Enterprises and research institutes deal with massive data volumes on a daily basis, also known as “big data,”² and require access to data that span centuries, also known as “long data.”³ 

Data availability has accelerated the rise of data-driven technologies, such as machine learning (ML),^4,5 a subclass of artificial intelligence technology, which has recently transformed several technological and scientific sectors, including nanophotonics.^6–13 However, the implementation of such data-intensive technologies demands significantly expanded data volumes with quick and reliable data access over long periods of time, posing a huge challenge to current data storage systems.

Today, data storage is based on magnetic storage technology, including hard disk drives (HDDs), solid-state drives, and tape. To store massive data volumes, data centers that use magnetic storage devices have been built all over the world. However, these devices only have terabyte-scale storage capacity, compel repeated data transfer to avoid losses, and consume a huge amount of energy.¹⁴ With the rise of “big data,” “long data,” and ML, there is a high demand for high-capacity, long-term data storage technologies.

Optical data storage (ODS), based on optical microscopy technology, promises low-energy consumption and high durability compared to magnetic storage technology.^15–18 However, expanding the storage capacity of existing ODS devices is critical. Nanophotonics advances have allowed for the engineering of nanoscale interactions between light and matter and the use of new physical dimensions of light and materials for optical data writing and reading. Such breakthroughs have paved the way for nanophotonics-enabled ODS devices with petabytes of data on a single optical disk (1 petabyte equals 1 thousand terabytes). However, the demand for high-capacity, long-term, robust, and reliable ODS necessitates breakthrough advancements in existing optical devices to permit future developments in artificial intelligence technology. New strategies for high-resolution, rapid, and reliable writing and readout methods facilitated by media with improved capabilities must be studied.

ML, which investigates computer algorithms capable of self-improvement via experience and data usage, has proven an unrivaled capacity to detect and anticipate data patterns and decode and extract information through image analysis.^19–22 Furthermore, ML has been combined with physical and chemical sciences to gain knowledge and create new fundamental principles and materials for nanophotonics applications.^23–26 Therefore, ML provides new approaches for high-resolution, accurate, fast, and robust optical data writing and reading, as well as the discovery, design, and optimization of nanomaterials and nanostructures with new functionalities for nanophotonics-enabled ODS applications. As a result, the combination of nanophotonics-enabled ODS with ML is expected to drive considerable innovation in the future generation of nanophotonics-enabled ODS.

In this Perspective, we review recent advancements in nanophotonics-enabled ODS and explore future prospects for its integration with ML in the creation of next-generation nanophotonics-enabled ODS. To begin, we examine the milestone developments in nanophotonics-enabled ODS, emphasizing the technological aspects that are critical to meeting data storage demands in the age of ML. Then, we outline emerging ML technologies that have the potential to disrupt writing and reading processes and the creation of new media for nanophotonics-enabled ODS. Finally, we will address future research possibilities and challenges in the field.

The challenge of storing ever-increasing amounts of data in a sustainable manner has resulted in ongoing technological revolutions in nanophotonics-enabled ODS toward higher storage capacity (Fig. 1). Current nanophotonics-enabled ODS methods have enabled petabyte-scale data storage capacity in a single optical disk and have even approached the storage capacity of the human brain.²⁷ 

In 2D ODS, optical data bits were written as 2D features (“pits”) in an optical disk’s flat area (“lands”). 2D ODS formats include CDs, DVDs, and Blu-ray Disks (BDs), with a storage capacity of up to 200 GB per disk and a lifetime of ∼30 years.²⁸ The BDXL format can store 100 GB on a triple-layer disk and 128 GB on a quad-layer disk. The Ultra HD Blu-ray format, also known as 4K Blu-ray, can store up to 100 GB of data on a triple-layer disk. Recently, a new generation of optical disks, known as Archival Disks (ADs), was released with a storage capacity of 300 GB per disk and a lifetime of up to 100 years.²⁹ Furthermore, future industrial developments are expected to increase AD storage capacity to 500 GB per disk and 1 TB per disk.²⁹ However, the storage capacity of these 2D formats is limited by the ability to optically write only one or a few data storage layers, while the volume of the optical disk remains unused. The main challenge is that focusing the writing laser beam into the ODS medium causes scattering loss. This issue was experimentally overcome by using the principle of two-photon (2P) absorption^30,31 to achieve 3D ODS with a storage density of >10¹² bits/cm³.^32,33 Because 2P absorption depends quadratically on the incident laser beam intensity, excitation was localized to a small 3D high-intensity region in the focal spot, which enabled us to reduce the separation between data storage layers without crosstalk. Recently, one-photon absorption systems have facilitated the writing of optical data bits in multiple layers by using illumination from modulated continuous-wave (CW) laser beams. A photothermal process was accounted for enabling a nonlinear response to achieve high-capacity 3D ODS.^34–36 

The need for a higher storage capacity compelled the invention of multi-dimensional ODS, in which multiple data states are multiplexed in the same location within an ODS medium. Polarization encoding was achieved based on polarization-sensitive refractive-index changes through 2P-induced isomerization of azo dye molecules in polymers for four-dimensional (4D) ODS with a storage capacity of ∼6 GB per disk.³⁷ Grayscale encoding in polymers was demonstrated for high-capacity 4D ODS with a storage capacity reaching 10 TB per disk.³⁸ 

Polarization and wavelength encoding was demonstrated through the photothermal reshaping of gold nanorods (GNRs) for five-dimensional (5D) ODS with a storage capacity of 1.6 TB per disk³⁹ [Fig. 2(a)]. The longitudinal surface plasmon resonance of GNRs is sensitive to wavelength and polarization, while the unique energy threshold necessary for photothermal reshaping offers axial selectivity for writing. By detecting the 2P luminescence mediated by longitudinal surface plasmon resonance, a non-destructive and crosstalk-free readout was achieved.

The development of high-capacity ODS with low-energy consumption is critical for sustainable big data storage. High-capacity, low-energy-consumption 5D ODS was demonstrated through polarization and wavelength encoding based on the generation of random hot spots in GNRs⁴⁰ [Fig. 2(b)]. In plasmonic systems, hot spots are highly localized modes with increased electromagnetic fields.^41–43 Plasmonic coupling strongly enhanced the 2P-induced absorption and luminescence of GNRs in the proximity of such hot spots. Multiplexing was achieved using the polarization, wavelength, and three spatial dimensions for 5D ODS with a storage capacity of 20 TB per disk and an energy consumption of 3.16 pJ/bit for writing and 0.42 pJ/bit for readout. Such results are equal to a 100-fold lower energy consumption compared to that of other diffraction-limited ODS techniques. Moreover, optical data writing with low-energy consumption enabled us to reduce the crosstalk between different writing channels for an improved quality of the stored optical data bits.

Secured and repeated optical data writing and reading over centuries is in high demand.^44,45 Intense industrial and research efforts have focused on the development of high-capacity ODS with a long lifetime. Commercial products have been released using inorganic ceramic materials for a lifetime of over 1000 years.⁴⁶ Permanent physical changes caused by laser irradiation, such as voids in polymers⁴⁷ and inorganic bulk materials with a high Young’s modulus (>50 GPa),⁴⁸ including glass and diamond,^49–52 provide an approach for a long lifetime and without data degradation.

Permanent 5D ODS was achieved through femtosecond laser nanostructuring of glass by controlling the laser beam polarization and intensity to manipulate the slow axis orientation and retardance of birefringence.⁵³ Because laser inscriptions on glass can resist to temperatures as high as 1000 °C, optical data bits can theoretically be preserved for thousands of years. However, these media have only enabled a storage capacity that is insufficient to meet today’s data storage demands, compelling the development of alternatives.

The combination of inorganic bulk materials with nanomaterials capable of advanced functionalities is a promising approach to increase the storage capacity in ODS media with a long lifetime. However, the large difference in synthesis temperatures between inorganic bulk materials and nanomaterials is a challenge for the realization of such optical devices. High-capacity, long-lifetime 5D ODS was demonstrated using nanoplasmonic hybrid glass composites⁵⁴ [Fig. 2(c)]. GNRs were incorporated into a hybrid glass composite through a sol–gel process at room temperature.⁵⁵ The inorganic phase of the host matrix increased the local Young’s modulus surrounding GNRs, which improved the lifetime of GNRs by minimizing undesired shape deterioration due to temperature perturbations. The control of the photothermal reshaping of GNRs enabled continuous multilevel optical data writing and readout with a storage capacity over 10 TB without changes in the baseline over 600 years.

Big data storage creates new challenges for data preservation technology, such as storage security. While encoding different physical dimensions of light through multi-dimensional ODS techniques enables us to boost the storage capacity, it also provides an approach for data encryption, which is a crucial component of data storage systems. Controlling the polarization of the writing beam allows data to be securely stored and retrieved using a polarization key. The flexibility of the encryption key can be obtained by rotating the polarization orientation, and multiple data states can be multiplexed. Through polarization-induced birefringence, data can be encrypted in two polarizations of the writing beam.³⁷ It was demonstrated that by tightly focusing a radially polarized beam, a longitudinal polarization state can be created within the focused region.^56,57

3D arbitrary polarization encryption was demonstrated using a single vectorial laser beam⁵⁸ [Fig. 2(d)]. Superimposing weighted radially and azimuthally polarized laser beams resulted in the optical configuration of the laser beam. Exploiting the polarization sensitivity of GNRs, 3D polarization encryption with five key states was demonstrated with potential for high-capacity ODS with high security.

Nanophotonic approaches enable us to encode data through multimode material responses for high storage capacity and high security. Graphene-based materials allow for changes in the refractive index and fluorescence intensity for hologram-encoded multimode recording and data access via 2P fluorescence readout.^59,60 The use of multimode recording by hologram encryption in conjunction with other physical dimensions, such as polarization, wavelength, and angular momentum, could result in high-capacity ODS with higher security. However, diffraction still limits the minimum achievable optical data bit size in multi-dimensional ODS, resulting in a limited storage capacity. Furthermore, the necessity for expensive high-intensity pulsed lasers, extremely sensitive polarization optics, and complicated manufacturing processes has hampered the usage of multi-dimensional ODS for commercial devices.

Recent progress in nanophotonics has enabled us to break through the diffraction-limit barrier and to shrink the size of optical data bits to the nanoscale for much-improved optical data bit density. Near-field super-resolution techniques, such as the use of near-field scanning optical microscopy probes,⁶¹ solid immersion lenses,⁶² and plasmonic lenses,⁶³ were first used to realize nanoscale ODS. However, these techniques exploit the properties of evanescent waves and thus can only operate close to the surface of ODS media, which limits their use to 2D applications. On the other hand, far-field super-resolution techniques, such as super-resolution optical microscopy⁶⁴ and optical nanolithography,⁶⁵ have laid the foundation for 3D nanoscale ODS toward petabyte-scale data storage capacity in a single optical disk.

Among these techniques, stimulated emission depletion (STED) microscopy has enabled optical imaging with nanometric resolution and shown the potential for nanoscale ODS with high storage capacity.^66,67 STED microscopy uses two spatially overlapped laser beams, one Gaussian-shaped laser beam for the activation of fluorophores and the other donut-shaped laser beam for deactivation, so that only fluorophores in the donut’s center fluoresce. STED microscopy improves the optical resolution by engineering the point spread function (PSF) by “ON” and “OFF” switching of the sample with the activation and deactivation laser beams and through the modulation of the intensity ratio of the two laser beams. The region defined by the donut’s center, where fluorophores can fluoresce, is described by the effective PSF of the STED microscope.^68,69 This is a Gaussian of full width half maximum d ≈ λ/[2NA(1 + I/I_s)^1/2], where λ is the wavelength of the donut-shaped laser beam, NA is the numerical aperture of the objective, I is the focal peak intensity at the donut-shaped laser beam crest, and I_s is the intensity that reduces the fluorescence ability by half. Importantly, STED microscopy allows for enhanced resolution to the size of single molecules (1–2 nm), which is theoretically possible for I > 10⁴I_s.

Nanoscale ODS through STED microscopy was demonstrated using a reversibly switchable enhanced green fluorescent protein (rsEGFP)⁷⁰ [Fig. 3(a)]. A 491-nm donut-shaped laser beam was used to turn the rsEGFP molecules to an “OFF” state, limiting the “ON” state to a region smaller than the diffraction limit in the donut’s center. Subsequently, a 532-nm Gaussian-shaped laser beam was used to make the “ON” state molecules permanent through photobleaching. Irradiation with a 405-nm laser beam switched the molecules from the “OFF” state back to an “ON” state, thus enabling subdiffraction optical writing of other features. Nanoscale optical data writing was achieved with a bit spacing of 250 nm for storing optical data bits with four times higher density than DVD storage density. Nanoscale optical bits could be readout five to ten times. Moreover, rsEGFP was used for repeated ODS in a scanning confocal setup, which demonstrates the potential for rewritable high-capacity ODS [Fig. 3(b)]. Irradiation with 405 and 491-nm laser beams enabled switching and reading optical data bits with a 50:1 on–off contrast. After 6600 cycles of writing and reading, the average fluorescence of the optical data bits was reduced by 35%, allowing rewritability for up to 15 000 read/write cycles.

Inspired by STED microscopy, an optical nanolithography technique known as “super-resolution photo-induction-inhibited nanolithography (SPIN)” has been developed for the writing of nanoscale features.⁷¹ The process of photoinhibition has been demonstrated through photoluminescence,^72,73 STED,⁷⁴ photodeactivation,^75,76 photoradical generation,^77–79 photoabsorption,⁸⁰ and photochromism and photoswitching.^81–83 SPIN has been used to achieve optical writing of 9-nm features in a 2P absorption polymer,^79,84 which potentially enables us to store petabytes of data in a single optical disk. 3D nanoscale optical writing was achieved using SPIN in polymers⁸⁵ [Fig. 3(c)]. An 800-nm Gaussian-shaped femtosecond pulsed laser beam was used for writing, and a 375-nm donut-shaped CW laser beam was used for inhibition. Optical data bits were written with a size of 80 nm and a spacing of 200 nm for a storage capacity of 30 TB per disk.

A challenge confronting current nanophotonics-enabled ODS based on bit-sequential scanning microscopy is low throughputs for data writing and readout, which are restricted to a few tens of Mbps.⁸⁶ It is essential to boost writing and readout throughputs along with the increasing storage capacity of ODS media. Optical parallelism through the generation of multifocal spot arrays using single laser shots is a promising approach for high-capacity ODS with high throughput. Multifocal spot arrays have been demonstrated through the use of diffractive optical elements,^87–89 microlens arrays,^90–92 and spatial light modulators (SLMs).^93–99 The use of an SLM is the preferred method due to its ability to dynamically update the intensity distributions in the focal plane by altering the incident phase patterns. However, the generation of uniform diffraction-limited multifocal spot arrays is challenging because of inaccurate phase patterns achieved through conventional fast Fourier transform methods.^100,101 The primary cause of this issue is that the depolarization effect in the focal region of a high-NA objective drastically changes the phase distribution.^102,103 In fact, for high-NA objectives (NA greater than 0.7), the paraxial approximation does not hold and effects such as apodization, depolarization, and aberration become pronounced and must be accounted for. The Debye theory (or Debye approximations) provides a diffraction integral that can be used to calculate the diffraction pattern of a high-NA objective. A method for phase retrieval based on Debye theory enables diffraction-limited multifocal spot arrays under high-NA focusing.⁹⁹ The intensity uniformity of diffraction-limited multifocal spot arrays was considerably enhanced, from 60% to 99% for high-NA objectives, by inserting a controlling factor in the phase generation process. Furthermore, the Debye-based Fourier transform method enables non-Airy,¹⁰⁴ cylindrically polarized,¹⁰⁵ aberration-free volumetric,¹⁰⁶ and polarization-multiplexed¹⁰⁷ multifocal spot arrays using high-NA objectives.

Parallelized SPIN was used for writing nanoscale optical data bits using dynamic and accurate phase modulation by SLMs [Fig. 3(d)]. The generation of a 10 × 10 multifocal array for both the photoinduction and photoinhibition laser beams enabled uniform 3D nanoscale optical data writing with a 100-fold improvement in data rates toward up to 4 Gbps. Furthermore, the Debye-based Fourier transform method can be integrated with other physical dimensions of laser beams,¹⁰⁵ holography,⁶⁰ or the use of incoherently superimposed orthogonal standing light waves^108,109 for throughputs of tens of Gbps. The combination of the Debye-based Fourier transform method with far-field super-resolution optical techniques enables for 3D nanoscale parallel optical writing and readout for high-capacity, high-throughput ODS.

Current SPIN methods typically necessitate high beam intensity to achieve initiation and inhibition, which results in high energy consumption and short lifetime. Moreover, typical materials for SPIN are not luminescent after writing, thus preventing optical data readout. These limitations demand for new optical medium platforms for nanoscale ODS with low-energy consumption and long lifetime.

High-capacity, low-energy-consumption nanoscale ODS was demonstrated in a new nanocomposite material based on lanthanide-doped upconversion nanoparticles (UCNPs) conjugated with graphene oxide (GO) flakes¹¹⁰ [Fig. 3(e)]. UCNPs can achieve photon upconversion, which is the optical conversion of low-energy near-infrared excitation photons into high-energy ultraviolet (UV) or visible emission photons.^111,112 Because of this feature, UCNPs have been used for many nanophotonic applications, including ODS.^113–116 Furthermore, UCNPs exhibit metastable excited energy levels with a lifetime of up to milliseconds,^111,114 which allows for low-power far-field super-resolution optical techniques, such as UCNP-enabled low-power STED microscopy.^117–119 GO is a graphene-like material that contains various oxygen functional groups.^120–126 The properties of GO can be tuned through reduction of the oxygen functional groups for nanophotonic applications, including ODS.^59,60,127

Nanoscale optical writing was achieved by using UCNPs to induce the nanoscale reduction of GO flakes through resonance energy transfer (RET).^128,129 Optical data bits were written in the nanocomposite with a size of ∼50 nm and a spacing of 195 nm using a 980-nm Gaussian-shaped CW writing laser beam and an 808-nm donut-shaped CW inhibition laser beam at intensities of 0.13 and 11.25 MW/cm², respectively. Optical data readout was achieved through fluorescence quenching microscopy^130,131 combined with UCNP-enabled STED microscopy. An estimated storage capacity of ∼700 TB per disk was achieved. By using UCNPs showing enhanced high-energy upconversion^132–136 and GO with optimized oxygen functional groups,¹³⁷ optical data writing can be achieved with an estimated energy consumption of <10 µJ/bit. In comparison, conventional ODS based on diffraction-limited methods, such as BD technology, requires picojoules of energy per bit.^15,16 On the other hand, nanoscale ODS based on super-resolution optical methods, such as SPIN technique, typically requires millijoules of energy per bit.^72–80 In addition, UCNP-enabled nanoscale ODS uses cheap CW lasers, minimizing cost compared to typical optical writing techniques that require high-priced and bulky high-intensity pulsed lasers. Thus, the use of UCNPs enables all-optical data storage at the nanoscale with low laser beam intensity for a reduction of the energy consumption and an increase in the lifetime.

Nanophotonics-enabled ODS has outperformed today’s magnetic data storage, which experiences a data density limitation. In the future, the combined use of far-field super-resolution optical techniques using nanomaterials and multi-dimensional multiplexing techniques might enable us to store hundreds of petabytes of data in a single optical disk.

Table I summarizes the milestone developments in nanophotonics-enabled ODS in terms of key technological parameters. The combination of these features in a single ODS medium could open up new avenues for the deployment of nanophotonics-enabled ODS devices that meet current data storage demands in the age of ML.

ML allows us to identify relationships in complex data that traditional statistical methods are unable to detect.⁴ Scientists who deal with “big data” and “long data” have routinely used ML to extract useful insights, predictions, and decisions from large-scale datasets. ML algorithms build an internal mathematical model for future data processing based on the training data and search for hidden relationships within the data. Over many data input cycles, the model fine-tunes its internal parameters until it converges to a specific optimized performance.

ML methods are broadly classified into supervised, unsupervised, and reinforcement learning.⁴ Supervised learning algorithms use inputs that have been labeled with a specific target performance. Unsupervised learning algorithms classify data using unlabeled inputs. Reinforcement learning algorithms interact with a dynamic environment in which they accomplish a specific task and learn through feedback. ML algorithms can handle high-dimensional inputs and detect complex and counterintuitive relationships. Once trained, they can generate new predictions and decisions almost immediately.

While breakthroughs in nanophotonics have enabled substantial improvements in nanophotonics-enabled ODS, the needs for high-capacity, long-term, robust, and reliable ODS necessitate breakthrough enhancements in existing optical techniques and media to enable future developments in artificial intelligence technology. ML has brought about new methods for robust optical data readout even beyond the diffraction limit, controlling adaptive optics (AO) technology to achieve aberration-free writing and inhibition beams for high-capacity ODS, and new pathways for the discovery, design, and optimization of nanomaterials and nanostructures for nanophotonics-enabled ODS applications. Section III examines emerging ML technologies that have the potential to significantly advance nanophotonics-enabled ODS for the future generation of nanophotonics-enabled ODS.

Optical data readout requires high resolution, high accuracy, and high throughput for quick and reliable data access. The ability to analyze optical images is thus crucial for optical data readout. A class of ML algorithms known as “deep learning” has recently proven useful in image processing applications^138,139 and has shown the potential to enable optical data readout with significantly improved resolution, accuracy, and throughput. Deep learning uses multilayered artificial neural networks (ANNs) to extract features and learn representations of data at various levels of abstraction.^140,141 ANNs are computer systems inspired by biological neural networks in the human brain that can be trained to address problems that would be challenging to solve with regular computer arithmetic. ANNs are built based on a network of interconnected elementary components known as “artificial neurons,” which are similar to biological neurons.

Deep learning based on a convolutional neural network (CNN) enabled us to improve the performance of optical microscopy,¹⁴² which has the potential for improving optical data readout in nanophotonics-enabled ODS applications. CNNs, a subset of deep neural networks that are deep-learning architectures based on ANNs with several stacked layers of artificial neurons, have been widely used in deep learning approaches to analyze image data.^138–141 The CNN was trained to statistically correlate a set of low-resolution optical microscopy images as input to high-resolution images as output [Fig. 4(a)]. The CNN was blindly tested on the images of biological specimens taken with a typical bright-field optical microscope at a low resolution. The CNN was fed low-resolution images acquired with a 40×/0.95NA objective lens and produced high-resolution images that matched the performance of a 100×/1.4NA objective lens, resulting in a diffraction-limited resolution increase from ∼0.36 to ∼0.28 µm. Furthermore, the network output showed a substantial improvement in spatial resolution on input images taken with a 100×/1.4NA objective lens, implying that it may also be used to achieve super-resolution images. The CNN extended the field of view by 6.25 times from 60 × 60 µm² per image from the high-NA objective lens to 150 × 150 µm² per image from the low-NA objective lens, which has the potential to increase the readout speed and thus data throughput. In addition, the network output exhibited a significantly enhanced depth of field. Such a deep learning-based imaging framework could be adapted to different optical microscopy modalities, which might be advantageous for higher-resolution, faster, and more accurate optical data reading over broad regions of ODS media.

Diffraction limits the size of the minimum achievable optical data bits and consequently storage density in ODS media. Alternative solutions for increasing storage density have been examined, including additional encoding dimensions, super-resolution optical techniques, and robust reading systems. Deep learning has enabled us to achieve imaging resolution enhancement,^143,144 fast super-resolution imaging,¹⁴⁵ and image reconstruction,¹⁴⁶ which offer the potential for quick and reliable subdiffraction optical data readout in nanophotonics-enabled ODS applications.

Deep learning was used for the subdiffraction optical readout of data bits that were stored in silicon nanostructures.¹⁴⁷ Silicon nanostructures show scattering with a high efficiency and tunability across the whole visible spectrum range due to the high refractive index of silicon.^148–150 Multiple data bits were encoded in the geometry of subdiffraction silicon nanostructures [Fig. 4(b)]. Subsequently, the subdiffraction optical readout of data bits was performed by utilizing a CNN to examine the scattering spectra of such structures [Fig. 4(c)]. The CNN was trained using the scattering spectra of hundreds of silicon nanostructures, each of which corresponded to a binary sequence. The trained CNN failed to interpret one single spectrum of the test data for 4 and 5 encoded bits, corresponding to 0.023% and 0.011% for the 4400 (4-bit) and 8800 (5-bit) test structures, respectively. The threshold bit error rate for practical use in a conventional ODS is 3.0 × 10⁻⁴. In comparison, a bit error rate of 1 × 10⁻⁵ has been achieved in an energy-gap-induced super-resolution read-only-memory disk in BD optics.¹⁵¹ Encoding up to nine data bits per diffraction-limited spot enables us to increase the storage density of BDs by 40%. The scattering intensity of a few wavelengths and the red, green, and blue color values from confocal dark-field optical microscopy images were used to achieve robust data retrieval with an accuracy of >99%, allowing for a simpler and parallel optical data readout method with an estimated throughput of up to 275 Mbps. This value is approximately eight times that of BDs, which yield 36 Mbps at 1× data rate.

However, such silicon nanostructures for data storage require fabrication by electron-beam lithography, followed by dry-etching and coating with protective layers, resulting in a time-consuming, complex, and energy-intensive process. In addition, because data bits were only encoded in planar arrays, this method was restricted to 2D applications in practice. These limitations necessitate the development of new techniques and materials for ML-enabled all-optical writing and readout at the nanometer scale to achieve high-capacity ODS.

Optical writing and reading data bits in multiple layers enable 3D ODS with higher storage capacity compared to 2D approaches. However, focusing on several data layers with different depths introduces aberrations that blur the focal spot, increase the volume of the written data bits, reduce the optical resolution of readout systems, and limit the achievable data layers. In optical data writing, aberrations can be corrected for by pre-shaping the light with an equal but opposite aberration.¹⁵² On the other hand, aberrations must be corrected in two paths during optical data reading because such a process involves passing light into the substrate to illuminate optical data bits and then returning the light out of the substrate.¹⁵³ 

AO technology has proven useful for aberration correction in 3D ODS based on the optical writing and readout of data bits as voids in multiple polymer layers.¹⁵⁴ AO systems typically include a method for measuring aberrations, an adaptive element for aberration correction such as deformable mirrors or SLMs, and a control system.¹⁵⁵ Only spherical aberration is detected in 3D ODS by perpendicular focusing through a refractive index mismatch. This effect is represented by Zernike polynomials, with the lowest-order spherical aberration mode prevailing. The performance of the system can be improved by removing this aberration mode. The depth-dependent spherical aberration correction needed for diffraction-limited performance in such 3D ODS application was achieved by maximizing a confocal optical microscopy reflection signal off the surface of different data layers. Moreover, AO technology has facilitated the production of high-quality donut-shaped laser beam profiles for the 3D STED microscopy imaging of samples with high aberration.¹⁵⁶ 

ML has been crucial in recent years for high-resolution optical microscopy enabled by AO technology,¹⁵⁷ which offers the possibility for revolutionary applications in high-capacity ODS.

ML-based AO technology was used for fast focusing of a Gaussian beam with wavefront aberration correction and recovery of near diffraction-limited focal spots¹⁵⁸ [Fig. 5(a)]. A CNN was trained to determine low-order aberrations from distorted PSF images detected using a CMOS camera with Zernike mode coefficients. Aberrations were corrected with an accuracy of about 90% utilizing compensation phase patterns loaded on a SLM at the back-pupil plane. The use of ML-enabled AO technology dramatically improved the PSF quality with an intensity increase by a factor of 3–5. The time required for phase reconstruction was less than 0.2 s and ∼0.08 s for compensation. This method was particularly advantageous since it eliminated the need for costly wavefront sensors to measure aberrations. Such a demonstration paves the way to achieve aberration-free laser beams for optical data writing and reading in multi-layer ODS applications. A similar ML-based AO approach allowed for restoring the focus of donut-shaped laser beams¹⁵⁹ [Fig. 5(b)]. This technique was used for the theoretical reconstruction of high-resolution images in scanning optical microscopy.

The integration of ML-based AO technology for the aberration correction of the Gaussian-shaped activation laser beam and donut-shaped deactivation laser beams could enable us to routinely perform efficient writing and reading of optical data bits with a size below 10 nm for the storage capacity of multiple petabytes of data per disk. ML for controlling AO technology and achieving aberration-free writing and inhibition beams offers enormous potential for improving nanoscale ODS based on super-resolution optical techniques toward higher storage capacity and lower energy consumption. Furthermore, the use of high-performance computers or workstations allows for improved speed for phase reconstruction to below milliseconds for higher throughput.^158,159 However, the refresh rate of the SLM restricts the correction speed in practice.

Recent advances in ML have enabled scientists to develop new materials with unique physical and chemical properties that have the potential to accelerate technological advancement in nanophotonics-enabled ODS.^23–26 ML has contributed to the creation of new strategies to inverse molecular and material design.^160–167 Typical inverse design problems require optimization in a high-dimensional space. In recent years, photonics research has focused on inverse design approaches, such as adjoint methods,¹⁶⁸ which mathematically reverse equations, and evolutionary algorithms,¹⁶⁹ which search the space step by step. Inverse design methods can be used to identify customized materials by starting with a specific desired performance. Such methods may enable us to develop new ODS media with targeted physical and chemical characteristics for implementation in nanophotonic techniques to achieve disruptive improvements in nanophotonics-enabled ODS applications.

ML was used to simulate light scattering and inverse design of photonic multilayer spherical nanoparticles¹⁷⁰ [Fig. 6(a)]. The nanoparticles comprised a silica core and alternating titanium oxide and silica shells. The thickness of different shells for such nanoparticles was fed into an ANN, which output the scattering cross section of the corresponding nanoparticles for different wavelengths. When applied to forward compute scattering spectra for which it had not been trained, the ANN was able to produce highly accurate simulations. The ANN was then used to solve a problem of inverse design using iterative back propagation. The ANN was successful in predicting nanoparticle structures that achieved a target scattering spectrum. Furthermore, the ANN could tackle complicated inverse design problems faster and more efficiently than numerical nonlinear optimization approaches. The ANN was also used as a tool for optimizing specific-wavelength or broadband scattering [Fig. 6(b)]. This method may be expanded to additional architectures, such as convolutions, dropouts, and residual networks, making it attractive for solving inverse design problems in a variety of contexts, including nanophotonics-enabled ODS applications.

The use of ML to develop new or improved materials with tailored physical and chemical characteristics has the potential to dramatically advance nanophotonics-enabled ODS. ML has demonstrated the potential to unraveling the intricacy of the data that links materials structures and characteristics for materials discovery and optimization.¹⁷¹ The quest for new media for nanophotonics-enabled ODS applications will be facilitated in the future by increasingly sophisticated and versatile computational methods powered by ML.

In this Perspective, we have reviewed recent milestone developments in nanophotonics-enabled ODS and described future research directions and challenges for its integration with emerging ML technologies to develop next-generation nanophotonics-enabled ODS.

When compared to magnetic data storage, ODS offers greater longevity and higher energy efficiency. As a result, there is still a huge opportunity for research and commercialization in developing new and improved ODS devices to store ever-increasing data volumes in the age of ML.

Scientists have been able to engineer the interaction of light and matter at the nanometer scale and the use of new physical dimensions of light and nanomaterials to achieve tremendous improvements in key technological parameters for nanophotonics-enabled ODS, such as storage capacity, energy consumption, lifetime, throughput, security, and rewritability. Embedding these unique features in a single ODS device may open up new possibilities for the implementation of energy-efficient optical data centers with exabyte capacity (1 exabyte equals 1 × 10⁶ terabytes).^15,16 Furthermore, the needs for high-capacity, long-term, robust, and reliable optical data storage necessitate breakthrough advances in existing optical devices to enable future developments of artificial intelligence technology.

ML, which allows a computer to “learn” through repeating operations, is quickly becoming crucial for nanophotonics research because it has proved to be an unequaled tool for detecting and forecasting data patterns and decoding and extracting information from images. ML has also been integrated with physical and chemical sciences to create new fundamental concepts and media. ML offers new pathways for innovations in nanophotonics-enabled ODS through robust optical data readout even beyond the diffraction limit, the control of AO technology to achieve aberration-free writing and inhibition beams for high-capacity ODS, and new pathways for the discovery, design, and optimization of nanomaterials and nanostructures for nanophotonics-enabled ODS applications.

The integration of nanophotonics-enabled ODS with emerging ML technologies promises new approaches for high-resolution, accurate, fast, and robust optical data writing and reading, as well as the discovery, design, and optimization of nanomaterials and nanostructures with new and upgraded functionalities for next-generation nanophotonics-enabled ODS. Along with the growing optical chip and optical fiber technologies, “intelligent” nanophotonics-enabled ODS integrated with ML will enable us to push a novel paradigm toward a sustainable optical large data storage platform.^15,16,172

M.G. acknowledges the funding support from the Zhangjiang National Innovation Demonstration Zone (Grant No. ZJ2019-ZD-005). Q.Z. acknowledges the funding support from the National Natural Science Foundation of China (Grant No. 61975123).

The authors have no conflicts to disclose.

All authors were involved in the writing of this paper.

Data sharing is not applicable to this article as no new data were created or analyzed in this study.