Artificial intelligence has been used to quickly and accurately model the 3D flow of light around arbitrarily shaped nanoparticles. Peter Wiecha and Otto Muskens at the University of Southampton in the UK demonstrated the modelling approach using a neural network that required just a single training procedure. Their technique could be used to design a wide range of optical devices that control the paths taken by light.
When light interacts with nanostructures that are smaller in size than the wavelength of the light, the result can be very different from how light interacts with larger structures and continuous media. The field of nanophotonics seeks to exploit this by designing nanoparticles with particular shapes and compositions with the aim of manipulating light in specific ways.
How light flows around such nanoparticles can be calculated using Maxwell’s equations of electromagnetism – at least in principle. In practice, however, the calculations can be very time consuming and it can take days to design and optimize complex structures.
Powerful tool
Artificial intelligence has recently emerged as a powerful tool for tackling optimization. Artificial neural networks can be taught to perform tasks through a knowledge of the basic rules underlying a system – and they have been used to approximate how spherical and H-shaped nanoparticles will interact with light. While successful, this technique can only be applied to simple, highly specific situations.
Wiecha and Muskens have taken a more generalized approach and have based their new technique on “convolutional” neural networks, which are commonly used for image analysis. The duo’s new system can quickly and accurately predict the 3D flow of light around nanoparticles with completely arbitrary shapes. A diverse variety of physical effects can be analysed with just a single training procedure – without the need to teach their neural network how to deal with numerous specific situations.
Computers ‘learn’ how nanoparticles scatter light
The researchers say that their approach could be applied to countless situations in nanophotonics. With further work, it could also be used for inverse design – whereby the required optical properties are input and the system designs the appropriate nanostrutures. Inverse design is currently extremely difficult to do and this capability could open a vast range of applications and areas of research that are unattainable today.
The duo’s neural network could soon enable researchers to monitor the performance of nanophotonic devices in real time, leading to more powerful physics experiments. Other applications could include computer chips with entirely optical components, nano-antennas which concentrate energy on molecular scales, and metasurfaces that can direct and control light.
Wiecha and Muskens are now aiming to improve the speed of their technique. They also hope to generalize the network even further to account for factors including multiple materials, arbitrary illumination, and larger geometries.
The research is described in Nano Letters.
