Pushing the limits of optical information storage using deep learning

I. SILICON NANOSTRUCTURES FOR DIGITAL INFORMATION ENCODING

• Each 2D array position encodes one bit, while an L-shaped frame distinguishes symmetric or rotational arrangements.A silicon block sets a bit to 1; its absence sets it to 0.Source: §I · p. 2
• For the first demonstration, structures encoded 2–5 bits and were fabricated by electron-beam lithography and dry etching of silicon-on-insulator substrates.The silicon layer was 90 nm high and the structures received a 190 nm SiO2 protective layer.Source: §I · p. 2
• Measurements included X- and Y-polarized dark-field spectra from 625 copies of each geometry.The 4-bit set comprised 20,000 spectra; the 2-, 3-, and 5-bit sets comprised 5,000, 10,000, and 40,000 spectra, respectively.Source: §I · p. 2, Figure 2 · p. 3
• Fabrication rastering produced 5–10% anisotropies between horizontal and vertical features, which the authors say can increase spectral uniqueness.The authors note that the asymmetries could also be corrected during mask design.Source: §I · p. 3

A. Read-out using scattering spectrum

• The readout approach trains an ANN on subsets of acquired spectra and tests it on spectra withheld from training.The optical scattering spectrum is used to distinguish structures and retrieve their bit sequences.Source: A. Read-out using scattering spectrum · p. 3

1. Network architecture

• The spectrum classifier combines a one-dimensional convolutional network with a fully connected layer and assigns each bit sequence an output neuron.The spectra enter through one or two channels, depending on whether one or both polarizations are used.Source: §1 · p. 3

2. Results

• For 2–5-bit datasets, quasi-error-free readout required at most around 60 training spectra per geometry, with lower bit densities requiring fewer.Figure 3 plots readout error against training samples per structure.Source: Figure 3 · p. 4
• 0.023% (4-bit) and 0.011% (5-bit) test error rates came from one misclassified spectrum each among 4,400 and 8,800 test structures.Quasi-error-free readout was reported for 2–5 bits using X, Y, or both polarizations.Source: §2 · p. 4
• Spectral variation was mainly associated with measurement conditions, especially confocal-hole position, and the authors report that the tested conditions still supported quasi-error-free operation.They note that acquisition stability could reduce this dispersion.Source: §2 · p. 4
• Short- and intermediate-wavelength windows as narrow as about 100 nm enabled error-free readout, while red-edge windows required up to about 200 nm for high accuracy.The authors attribute the red-edge limitation to fewer distinguishing features and no resonances above about 750 nm in the tested geometries.Source: §2 · p. 4
• Three sampled wavelengths yielded readout accuracy above 99%, while four or five reduced worst-digit error below 0.5%.The reported worst-digit error with three wavelengths was still around 10%.Source: §2 · p. 5

B. Parallel read-out using RGB color from microscope images

• RGB values from conventional dark-field microscopy images were investigated as a simpler and faster data-acquisition scheme.The approach follows the finding that reduced spectral information can suffice for retrieval.Source: B. Parallel read-out using RGB color… · p. 5

1. Training data and network architecture

• The training data used average RGB values from each structure, normalized so R+G+B = 1, with scattered intensity stored separately.Four datasets used X polarization, Y polarization, both polarizations, or both polarizations plus intensity.Source: §1 · p. 5
• An entirely fully connected neural network was used to classify the RGB-based datasets.The figure shows representative 3×3 arrays for 4-bit geometries under X- and Y-polarized dark-field imaging.Source: §1 · p. 5

2. Results

• Single-polarization RGB datasets showed partial mixing and insufficient identification, with some individual-digit error rates well above 50%.The corresponding t-SNE plots showed clearer separation only when both polarizations were used.Source: §2 · pp. 5–6
• Below 1% average error for 2–4 bits and below 3% for 5 bits were reported when both polarizations and brightness were used.Combining both polarizations without brightness reduced average error below 10%; some digits with similar color and brightness remained difficult to distinguish.Source: §2 · p. 6
• RGB imaging can capture many thousands of structures in one large-area measurement, supporting massively parallel readout.The passage also describes smartphone-based dark-field microscopy as a possible future low-cost option.Source: §2 · p. 6

1. Training data and network architecture

• To manage 512 geometries, four measured copies per geometry were expanded into 300 random spectra with intensity deviations up to ±10%.Source: §1 · p. 7

2. Results

• Far below 1% readout error is achieved with 4 or 5 probed wavelengths for 9-bit decoding.Three probed wavelengths yield approximately 2% error.Source: §2 · p. 7
• Significant errors begin around 15% simulated noise; at 20%, average error is 4.16%, but worst-digit error reaches 35%.Source: §2 · p. 7
• With a 10% spectral-deviation tolerance, 9-bit structures yield around 40% higher information density than Blu-ray; 4-bit structures reach 75%.Source: §2 · p. 7, Figure 6 · p. 8

• Deep neural networks retrieve optical information encoded in photonic-nanostructure geometry, with demonstrated structures encoding up to 9 bits per diffraction-limited area.Source: §IV · p. 8
• A few discrete wavelengths or RGB values from dark-field microscopy images provide precise readout; the RGB route may use simple, inexpensive equipment and enable massively parallel retrieval.Source: §IV · p. 9
• The approach is presented as applicable to high-density planar optical storage and other nano-optics classification tasks, including plasmonics and biological-specimen identification.Source: §IV · p. 9

A. Nanofabrication of planar Si structures

• Silicon-on-insulator structures were patterned by electron-beam lithography, transferred by reactive-ion etching, then planarized with HSQ and annealed to form SiOx.Source: A. Nanofabrication of planar Si… · p. 10

B. Confocal dark-field microscopy

• Scattering spectra were collected using a confocal dark-field microscope, while microscopy images were captured with a color CCD and each structure's average RGB value was measured.The main-text spectra used 0.2 s acquisition; a 0.5 s dataset produced similar results.Source: B. Confocal dark-field microscopy · pp. 10–11

C. Expansion of 9 bit experimental training data set

• Because acquiring spectra from many copies of all 512 geometries was impractical, four measured copies per geometry were combined into synthetic spectra using random weights summing to one.Source: C. Expansion of 9 bit experimental… · p. 11
• The expanded dataset applied random intensity scaling from 0.9 to 1.1 and generated 300 semi-experimental spectra for neural-network training.Source: C. Expansion of 9 bit experimental… · p. 11

Simulated dataset

• All 512 possible 9-bit geometries are simulated under X- and Y-polarized illumination, with randomized spectral perturbations used to vary noise from 10% to 25%.The geometry has 3 × 3 silicon blocks, each 105×105 nm2 and 90 nm high, with a 45 nm-wide L-shaped border.Source: Simulated dataset · p. 11

Green Dyadic Method

• The Green Dyadic Method discretizes the nanostructure into cubic meshpoints with dipolar responses, producing 3N coupled equations for the fields.The structures are modeled in a homogeneous SiO2 environment 30 nm above a silicon substrate.Source: Green Dyadic Method · p. 11
• The Green’s Dyad combines vacuum and substrate terms, and its divergence at coincident meshpoints is handled with a normalization scheme.The cubic mesh uses cell volume vi = d3.Source: Green Dyadic Method · p. 11
• Inverting equation (2) with LU decomposition yields the internal near-field used to calculate scattering cross-sections.The GDM is a frequency-domain method based on a generalized propagator.Source: Green Dyadic Method · p. 11

• The artificial neural networks are implemented in Python using TensorFlow.Source: E. Training artificial neural networks… · p. 12

• Spectra are median-filtered, smoothed, and downsampled to 99 wavelengths before training.The median filter uses a 6 nm kernel, the moving-average smoothing a 15 nm kernel, and downsampling an order 8 Chebyshev type I filter.Source: Digit retrieval using scattering… · p. 12
• The retrieval model uses three convolutional layers followed by two fully connected layers, with max pooling, batch normalization, and regularization.The convolutional layers use leaky ReLU; the fully connected layers use tanh, L2 regularization is applied to the convolutional part, and dropout with 80% keep probability to the fully connected part.Source: Digit retrieval using scattering… · p. 12
• Training uses Adam with batch size 64, categorical cross-entropy, learning rate 0.0001, and 200 epochs, or 2000 epochs for 9-bit structures.Source: Digit retrieval using scattering… · p. 12

• For discrete-wavelength input, intensities are averaged across three neighboring values to reduce noise.Source: Discrete wavelengths and RGB digit… · p. 12
• Five-wavelength retrieval is designed to probe spectral regions carrying the most information, with the longest wavelength limited to about 740 nm.The red spectral region was found to contain very little information.Source: Discrete wavelengths and RGB digit… · p. 12
• For RGB retrieval, pixels at least three times brighter than background are grouped per structure, and their average RGB values are normalized to R + G + B = 1.The scattered intensity is stored separately as brightness.Source: Discrete wavelengths and RGB digit… · p. 12

Network architecture

• The RGB and discrete-wavelength datasets are classified by a three-layer fully connected network with a 16-neuron softmax output.The input consists of wavelength intensities or RGB values, optionally including intensity; training uses 300 samples per geometry.Source: Network architecture · p. 12

Multi-wavelength based readout

• A multi-laser reader could retrieve 9 bits per passage, comparable in performance to a disc reader that retrieves a single bit.The proposed system would probe scattering intensity at different wavelengths sequentially.Source: Multi-wavelength based readout · p. 12

RGB based readout

• 20 images per second is the current exposure-limited rate, about five times below the estimated 100 images per second needed to match disc performance.The estimate assumes approximately 80,000 structures per image, each encoding 5 bits.Source: RGB based readout · p. 13
• 275 Mbit/s is the projected readout rate with a 2.5 × 2.5 larger captured area and a 0.01 s exposure.The passage describes the exposure time as a conservative assumption.Source: RGB based readout · p. 13