Machine learning helps improving photonic applications
The computer simulation shows how the electromagnetic field is distributed in the silicon layer with hole pattern after excitation with a laser. Here, stripes with local field maxima are formed, so that quantum dots shine particularly strongly. Picture. C. Barth/HZB
Photonic nanostructures can be used for many applications, not just in solar cells, but also in optical sensors for cancer markers or other biomolecules, for example. A team at HZB using computer simulations and machine learning has now shown how the design of such nanostructures can be selectively optimised. The results are published in Communications Physics.
Nanostructures can increase the sensitivity of optical sensors enormously – provided that the geometry meets certain conditions and matches the wavelength of the incident light. This is because the electromagnetic field of light can be greatly amplified or reduced by the local nanostructure. The HZB Young Investigator Group “Nano-SIPPE” headed by Prof. Christiane Becker is working to develop these kinds of nanostructures. Computer simulations are an important tool for this. Dr. Carlo Barth from the Nano-SIPPE team has now identified the most important patterns of field distribution in a nanostructure using machine learning, and has thereby explained the experimental findings very well for the first time.
Quantum dots on nanostructures
The photonic nanostructures examined in this paper consist of a silicon layer with a regular hole pattern coated with what are referred to as quantum dots made of lead sulphide. Excited with a laser, the quantum dots close to local field amplifications emit much more light than on an unordered surface. This makes it possible to empirically demonstrate how the laser light interacts with the nanostructure.
Ten different patterns discovered by machine learning
In order to systematically record what happens when individual parameters of the nanostructure change, Barth calculates the three-dimensional electric field distribution for each parameter set using software developed at the Zuse Institute Berlin. Barth then had these enormous amounts of data analyzed by other computer programs based on machine learning. “The computer has searched through the approximately 45,000 data records and grouped them into about ten different patterns”, he explains. Finally, Barth and Becker succeeded in identifying three basic patterns among them in which the fields are amplified in various specific areas of the nanoholes.
Outlook: Detection of single molecules, e.g. cancer markers
This allows photonic crystal membranes based on excitation amplification to be optimised for virtually any application. This is because some biomolecules accumulate preferentially along the hole edges, for example, while others prefer the plateaus between the holes, depending on the application. With the correct geometry and the right excitation by light, the maximum electric field amplification can be generated exactly at the attachment sites of the desired molecules. This would increase the sensitivity of optical sensors for cancer markers to the level of individual molecules, for example.
The software used as well as the data can be downloaded free.
Published in Communications Physics (2018). “Machine learning classification for field distributions of photonic modes”, Carlo Barth & Christiane Becker
DOI:10.1038/s42005-018-0060-1
- Catalysis research at HZB gets new facilityAs part of the CatLab project, HZB has acquired a unique facility for measuring the catalytic performance of thin-film catalysts. Built by ILS in Adlershof, it has now been delivered. The facility consists of a total of eight chemical reactors in which catalytic systems can be tested. At over €2.5 million, this is the largest single investment in the CatLab project.
- Protein crystallography at BESSY II: faster, better and more and more automaticMany diseases are linked to malfunctions of proteins in the organism. The three-dimensional architecture of these molecules is often highly complex, but it can provide valuable insights into biological processes and the development of drugs. X-ray diffraction at the MX beamlines of BESSY II can be used to decipher the 3D structure of proteins. To date, more than 5000 structures have been solved at the three MX beamlines. Here, we present a review and an outlook with Manfred Weiss, head of the research group for macromolecular crystallography.
- Humboldt-Fellow at HZB-Institute for Solar Fuels: Alexander R. UhlAlexander R. Uhl, UBC Okanagan School of Engineering in Kelowna, Canada, aims to develop with Roel van de Krol from the HZB Institute for Solar Fuels an efficient and inexpensive photoelectrolyser for producing hydrogen using sunlight. His stay is being funded by the Alexander von Humboldt Foundation.