Cerabyte–CERN Collaboration: Optical Data Storage and QR Code Decoding

Cerabyte is developing ultra-long-term optical data storage on ceramic-coated glass media. CERN’s scientific programme generates exabytes of data, requiring archival systems that are not only scalable but also sustainable for centuries. Cerabyte’s approach encodes data as matrices of QR codes written by femtosecond-laser (fs) ablation into durable ceramic layers and read back by high-speed optical microscopy. The writer and reader are therefore different technologies (fs-laser vs microscope), and unlike magnetic tape the ceramic medium is write-once, not erasable.
In this project, multiple approaches were investigated to achieve industrial-scale decoding throughput (on the order of 500 FPS, corresponding to 10.000 −- 30.000 QR codes/s). Initial experiments with deep learning detectors (e.g., YOLOv8) showed limitations in both accuracy and throughput for dense layouts. The focus therefore shifted to classical computer vision to localize 4×4 QR matrices and extract per-code crops. Several decoding libraries were benchmarked; OpenCV’s WeChat QR decoder emerged as the most effective balance of robustness and speed. Through multithreaded CPU implementation, the pipeline achieved sub-5 ms per QR and exceeded 75% decoding accuracy on real microscope samples despite imperfections in the medium. Further acceleration using GPU inference with OpenCV DNN increased throughput to approximately 18.000 QR/s. A hybrid approach that combines classical computer vision with lightweight deep learning refinements out-performed either method alone.
The contributions include: (i) a repeatable imaging and decoding workflow for prototype carriers, (ii) identification of limitations of deep learning for dense QR scenarios, (iii) a scalable computer-vision-based solution with validated benchmarks, and (iv) a roadmap toward a fully GPU-first, production-grade pipeline.