Scientists propose using deep learning for error correction when reading optical storage based on silicon nanostructures. This method allows for significantly higher density optical storage.
Researchers have been working on successors for DVD and Blu-ray with higher storage density for some time now. With those existing optical media, bits are stored on a surface with a size related to the diffraction limit. That is the distance between two adjacent points that the laser can distinguish.
Many initiatives to increase storage density have drawbacks, write researchers at the University of Toulouse. They do not offer much improvement over existing optical storage techniques or are complex in the storage medium or readout system.
Photonic nanostructures hold promise, the researchers say, as a tool to control light at the nanoscale. They acknowledge that this technique also has drawbacks, because noise and small deviations in the geometry, such as due to defects, can lead to errors in the optical detection. In addition, the material used, often gold, does not make production scalable.
The researchers therefore propose to use silicon nanostructures. These support low-loss optical resonance and are tunable across the entire visible spectrum. In addition, the production is proven and cheap thanks to cmos technology.
The geometry of the structures is a layout of blocks and ‘holes’ that successively represent the 1 and the 0. For example, two blocks and two ‘holes’ involve four bits and variants of two, three or five bits are also possible. The structure is delimited by an L-shaped wall and the blocks have a size of 120x120nm with a height of 90nm.
Darkfield microscopy makes it possible to map the spectrum of the scattered light reflected by the various structures. To obtain the bit sequences that represent the structures, the French scientists had a neural network train on a dataset of acquired spectra. They used an architecture optimized for pattern recognition.
The results were ‘virtually’ error-free and measuring a few wavelengths instead of the entire spectrum proved sufficient to read the information. The technique can be simplified even more by focusing on the measurement of rgb values that can simultaneously read thousands of structures on a large surface with a single measurement, the scientists argue.
The researchers do not report anything about the speed of the read or write, nor is it clear how much stretch there is to further increase the storage density. They do, however, point to the advantage of being able to read massively in parallel and the possibility that reading can be done with simple and cheap systems, which could make the research a starting point for a new storage technique.
They published their research in an article called Pushing the limits of optical information storage using deep learning in Nature Nanotechnology. The entire article is only available for a fee, but a previous publication can be viewed freely.