Tel Aviv University researchers have developed a diffusion model-based method that designs flat optical components in minutes — instead of hours and days — and demonstrates high accuracy in tasks such as beam splitting and polarization separation; published in ACS Photonics
Researchers from the School of Electrical and Computer Engineering at Tel Aviv University have developed a groundbreaking method for designing tiny optical components using artificial intelligence. The method allows flat optical components – known as metasurfaces – to be designed in just a few minutes, instead of hours or even days as was the case until now. This is a significant leap forward in the field of optics, with the potential to change the way cameras, sensors, and augmented reality systems are developed.
The research was conducted by research students Liav Chen and Erez Yosef, under the supervision of researchers Prof. Raja Giris, Prof. Dan Raviv and Prof. Kobi Shuyer, all from the School of Electrical and Computer Engineering at Tel Aviv University. The research was published in the scientific journal ACS Photonics
The research team explains that in recent decades, the world of optics has undergone a dramatic change: Instead of thick, heavy lenses and optical components, researchers are developing “metasurfaces” — ultra-thin structures a few hundred nanometers (millionths of a millimeter) thick, made up of tiny structures called metaatoms. Metasurfaces are able to control the direction, intensity, and polarization of light, thereby performing operations that previously required large, expensive components.
Designing a metasurface is a particularly complex engineering task. It is an “inverse design” problem – where you know how you want the light to behave, but you don’t know what the physical structure that will make it happen should look like. Until now, solving the problem has required lengthy simulations that sometimes lasted for days.
In the new study, researchers at Tel Aviv University managed to dramatically shorten the process using a diffusion model – an advanced type of generative neural network (Generative AI), similar to models that create images, but here it is used to design tiny optical structures.
The researchers created a vast database of examples mapping the structure of a metasurface to the light scattering pattern it creates. The model learned these complex relationships, and then was able to generate new designs in record time – less than 30 minutes – and with a very high level of accuracy.
The researchers demonstrated the effectiveness of the method on a variety of optical tasks, including designing a metasurface that splits a light beam into a number of equal directions, as well as a component that separates horizontally polarized light from vertically polarized light – an important function in advanced optical systems.
Furthermore, the developed method is flexible and adaptable to new tasks, different types of materials, and diverse physical conditions – thanks to a unique mechanism for building high-quality datasets for training the model. According to the researchers, the new method illustrates how generative artificial intelligence – a technology primarily associated with creating art and images – can become a powerful scientific and engineering tool. In the future, approaches of this type may enable real-time design of customized lenses and sensors, streamline manufacturing processes, and contribute to the development of new technologies in the fields of medicine, communications, and wearable electronics.
More of the topic in Hayadan:
2 תגובות
There is a large gap between what is written about the optical technologies available in the world today and reality.
Wonderful, now they'll invent lenses for glasses that repel dirt.