Laser writing in glass for dense, fast and efficient archival data storage
Summary
Microsoft Research’s Project Silica team demonstrates Silica: a full archival storage system based on femtosecond laser direct writing in glass. The platform writes 3D arrangements of small volume pixels (voxels) inside durable glass using two complementary regimes — birefringent voxels (pseudo-single-pulse) and phase voxels (single-pulse) — combined with closed-loop emission control, machine-learning decoding and error-correction to deliver reliable, automated writes and reads at high density and throughput. Accelerated ageing tests extrapolate data lifetimes beyond 10,000 years at room temperature, and the system is shown to meet the main metrics required for production archival storage.
Key Points
- Silica uses femtosecond laser direct writing to store data inside glass platters, providing thermal, chemical and electromagnetic resilience compared with conventional media.
- Two voxel regimes: pseudo-single-pulse birefringent voxels (higher density) and single-pulse phase voxels (broader material compatibility).
- Headline metrics: birefringent voxels achieved 1.59 Gbit mm−3 (usable ≈4.84 TB/platter) and 25.6 Mbit s−1 write throughput; phase voxels achieved 0.678 Gbit mm−3 (≈2.02 TB/platter) and 18.4 Mbit s−1; a 4-beam multibeam setup reached 65.9 Mbit s−1.
- Closed-loop emissions monitoring stabilises pulse energy; convolutional neural networks decode images into symbols and LDPC + erasure coding ensure data integrity.
- Automated write/read pipeline demonstrated on billions of voxels; accelerated thermal annealing experiments project >10,000-year voxel lifetime at room temperature.
Content summary
Silica receives a user bitstream, applies forward error correction and maps grouped bits into symbols that are encoded into voxels inside 2 mm glass platters. Voxels are written layer-by-layer using polarisation or amplitude modulation, with precise focusing and polygon scanning to reach sub-micron lateral pitch and micrometre layer spacing. The team developed two efficient writing regimes that minimise pulses per voxel and therefore maximise throughput and energy efficiency. Reading uses wide-field microscopy (polarisation-resolved for birefringent voxels, Zernike phase-contrast for phase voxels) and a CNN-based inference pipeline that outputs symbol probabilities; soft-decision LDPC decoding and erasure codes recover user data.
The paper includes a full system analysis: voxel quality, usable capacity per 120 mm×120 mm×2 mm platter, write efficiency (nJ per bit), and throughput. It also describes symbol-selection optimisation, closed-loop emission control to flatten spatial/temporal variations, and multibeam scaling. Practical engineering choices and trade-offs (material choice, NA, beam count, camera/read FOV) are explored, and routes to scale throughput and reduce write energy are outlined.
Context and relevance
Global data volumes are doubling roughly every three years and long-term archives (legal, cultural, research) need media that don’t require frequent migration. Current magnetic and solid-state archival media degrade in years or decades and demand costly refresh cycles. Silica targets that problem by using glass — intrinsically stable against moisture, temperature swings and electromagnetic interference — and demonstrating a whole-system approach that meets archival metrics (density, durability, integrity, throughput). The work therefore matters to cloud providers, national archives, institutions keeping long-term records and anyone planning storage with century-to-millennia horizons.
Why should I read this?
Want to know what could replace tape and hard drives for ultra‑long archives? This paper actually builds and tests the whole pipeline — writing, reading, decoding and ageing — so you won’t need to sift through fragmented lab demos. If you care about future‑proof, low‑maintenance storage (or you design storage systems), it’s worth a skim — and a deeper read if you’re planning to adopt or develop archival hardware or software around glass media.
Author style
Punchy: the team doesn’t just tinker with voxels — they deliver a production‑minded stack, quantify all key metrics and demonstrate repeatability at scale. If archival storage is in your remit, this is not an academic curiosity: read the details to understand real trade‑offs (materials, NA, beam parallisation, read hardware and FEC) you’ll need to address to move beyond prototypes.
Source
Article date: 18 February 2026