Introduction to Silica Glass Data Storage

The Silica storage system is meticulously evaluated using several key metrics, specifically focusing on its advanced data storage capabilities and operational efficiency. This innovative system stores user data within microscopic voxels embedded in glass, where the quality and density of this storage are absolutely paramount for its overall performance.

Key Performance Metrics

The system's efficiency and capability are rigorously assessed through the following performance indicators:

• Voxel Quality (Q): This metric represents the average bits per voxel, derived from the total user bits (B) across a given number of voxels (n V). It inherently remains less than the theoretical maximum possible bits per voxel (Q w) due to the inclusion of redundant bits vital for error correction. A crucial quality factor (q = Q / Q w) is subsequently defined.
• Data Density (ρ): Measured in Gbit mm⁻³, data density quantifies the voxel quality stored per unit volume of glass. It is directly derived from Q and the effective volume (V) of glass actively utilized for data storage.
• Usable Capacity: This metric specifies the total user bits that can be reliably stored within a standard 120 mm square, 2 mm thick glass platter. It meticulously accounts for all engineering overheads, providing a practical measure in TB per platter.
• Write Throughput (θ): Representing the speed at which data is written to the glass, throughput is defined as the peak rate in bit s⁻¹. It is directly influenced by the laser's repetition rate and the number of active beam lines.
• Write Efficiency (η): This metric assesses the energy consumed per user bit written, expressed in nJ per bit. A lower write efficiency signifies superior performance, as it allows for greater parallelism in writing operations using the same amount of laser energy.
• Lifetime: An experimental estimation of the duration for which data reliably remains stored and readable within the glass medium.

Data Writing Process

The writing system leverages an amplified femtosecond laser, operating at a 516 nm wavelength with a pulse duration ranging from 300 to 1000 fs and a repetition rate of 10 MHz. These precisely controlled laser pulses are routed through a tunable attenuator and then modulated according to the specific type of voxel being written.

System Components

A sophisticated setup facilitates the writing process:

• A self-air-bearing polygon scanner, rotating at 10,000–50,000 rpm, accurately directs scanned laser pulses.
• These pulses pass through a custom f-theta scan lens and a relay lens.
• An objective lens focuses the pulses precisely inside a 2 mm thick glass platter, which moves on an xy translation stage.
• Writing depth can be finely adjusted by manipulating the objective lens and its dedicated spherical aberration correction collar.

Voxel Types

The system supports two distinct types of voxels:

• Birefringent Voxels: These are formed in fused silica glass (Heraeus, Spectrosil 2000). Their creation necessitates a tunable beam splitter to generate separate seed and data pulses, followed by a polarization modulator utilizing Pockels cells with elliptical polarization to induce the required modifications within the glass.
• Phase Voxels: Written into borosilicate glass (Schott, BOROFLOAT 33), this process primarily employs an acousto-optic modulator (AOM) for amplitude modulation. Multibeam writing is also integrated, where a single laser source is split into four distinct beams, each independently modulated and focused.

Emissions-Based Control of Voxel Writing

During the critical phase of voxel formation, a dense, superheated plasma is generated, emitting white light. This characteristic emission is ingeniously collected by the write objective and forms the basis for a closed-loop control system designed to stabilize the writing process.

Mechanism and Control Aspects

A CMOS sensor meticulously captures photoemission images during the writing operation. This data is then used for two key control functions:

• Offline Flattening: This initial calibration step compensates for static inconsistencies across the scanning space, such as variations in polygon facet reflectivity. It achieves this by precisely mapping the AOM modulation to the observed emission levels.
• Closed-Loop Control: This dynamic process continuously adjusts the laser power via the AOM to maintain a pre-defined emission target. It effectively compensates for real-time variations like temperature fluctuations, ensuring consistent voxel quality. This control mechanism is reset for each new layer, and the camera shutter time is configured to integrate over at least one complete polygon revolution for stable measurement.

Data Reading Hardware and Process

Both birefringent and phase voxels are read using a sophisticated wide-field microscope equipped with an sCMOS camera and LED illumination. Essential fiducial markers, previously written into the glass, are critical for calibrating the sample position, enabling precise automated tracking. The system intelligently identifies optimal in-focus positions by diligently monitoring a variance-based sharpness metric as the sample navigates along the z-axis.

Birefringent Read

• A custom wide-field polarization microscope, utilizing a 525 nm LED, is employed.
• The sample is illuminated with circularly polarized light.
• Detection involves two liquid crystal variable retarders, which precisely set the required polarization detection states.
• Three distinct polarization states are strategically arranged at 120° intervals around the Poincaré sphere, optimizing the signal processing for robust data retrieval.

Phase Read

• This method utilizes a custom Zernike phase-contrast microscope illuminated by a 445 nm LED.
• While providing clear voxel visibility, this technique inherently exhibits poor optical sectioning capabilities.
• To mitigate this limitation and reduce the bit error rate, two images are captured per data layer: one at maximal voxel contrast and a second image several micrometers deeper, where voxel contrast is intentionally reduced and inverted.

Lifetime Conditions

The long-term durability of phase voxels is critically evaluated through macroscopic measurements that diligently track thermal erasure in annealed glass platters. The diffraction efficiency of a 405 nm laser beam passing through the modified glass serves as a reliable proxy for assessing changes in refractive index modulation. Samples undergo annealing at elevated temperatures, ranging from 440–500 °C. The subsequent decay curves of diffraction efficiency are meticulously fitted to a stretched exponential model, allowing for the determination of an activation energy of 3.28 eV. This activation energy is a crucial factor, enabling the estimation of the data's projected lifetime.

Data Decoding and Error Correction

Advanced machine learning techniques, specifically Convolutional Neural Networks (CNNs), are central to inferring symbols from captured sector images. The comprehensive decoding pipeline encompasses several stages: initial pre-processing, accurate symbol inference, precise symbol-to-bit mapping, and robust error correction.

Machine Learning Model

• A Convolutional Neural Network (CNN), pre-trained on known data patterns, receives stacked images (which include crucial context images for phase voxels) as its input.
• Its output is a 2D array of symbol probabilities for each individual voxel.
• This sophisticated model effectively accounts for complex spatial information, optical scattering effects, and potential interference, ensuring high-fidelity symbol recognition.

Redundancy Optimization and Error Correction

The ultimate goal is to maximize the product of the LDPC code rate and the erasure code rate, thereby achieving the highest possible useful data density.

• LDPC Codes: Employing the robust 5G wireless telephony standard, Low-Density Parity-Check (LDPC) codes are utilized for intricate error correction within individual sectors. They benefit from soft-decision decoding, which leverages the bit probabilities provided directly by the CNN.
• Erasure Codes: These powerful codes strategically spread k sectors of user data across n sectors. This ingenious design allows for the complete recovery of data as long as a minimum of k out of n sectors are perfectly decoded, significantly enhancing data resilience.
• Density Maximization: The overarching objective of the system is to optimize the product of the LDPC code rate and the erasure code rate, which directly translates to achieving the highest possible useful data density.
• Extending Binary Gray Codes: To further enhance channel capacity, the system explores and implements non-powers-of-two symbol counts for binary encoding. Crucially, it maintains a "Gray property," where confusing adjacent symbols results in the flip of only a single bit, minimizing error propagation.
• Symbol Selection Optimization: The system meticulously optimizes both the number of symbols and their corresponding energy modulation levels. A regression model is employed to predict the optimal modulation for each voxel. Subsequently, an sophisticated optimization process intelligently selects an optimal subset of N modulations from a larger pool of M candidates (e.g., 31 options) to maximize the mutual information between the decoded and the original transmitted data.