Deep learning network demonstrates ground-breaking understanding of light-matter interaction
Physicists at the University of Southampton have taught an artificial neural network to develop an intuitive understanding of the flow of light in nanostructures.
The pioneering study opens new routes for nano-optical modelling at unprecedented speed, potentially enabling formerly impossible real-time performance of nanophotonic devices found in optical chips, metasurface flat lenses and quantum technology.
These applications offer optical properties 'on demand' but typically need to be designed in a process that requires time-consuming numerical simulations, taking hours or even days in complex systems. The new research, published this month in Nano Letters, demonstrates speeds that are thousands of times faster by making use of neural networks.
Lead author Dr Peter Wiecha, of Physics and Astronomy, says: "Deep learning neural networks can be trained to solve a wide range of situations by extracting some basic underlying rules. Here, we show for the first time that a neural network can be used to infer the precise flow of light at the nanoscale. This three-dimensional solving network is a ground-breaking demonstration of what deep learning can be capable to do in a scientific context."
The interactions of light with nanostructures are governed by a set of electromagnetic equations that are generally applicable from radio waves to light waves. These Maxwell's equations represent some of the most profound foundations of physics.
In their new work, the researchers showed that neural networks can be developed to learn the essential light-matter interactions resulting from Maxwell's equations and hence capture a much more generalised class of problems related to the electromagnetic fields inside the nano-optical devices. Knowledge of these fields in all three dimensions allowed researchers to reconstruct most of the characteristics of the light-matter interaction without any approximations.
Co-author Professor Otto Muskens, Leader of the Integrated Nanophotonics Group, says: "Manifold applications will arise from this work now we have proven that a neural network can learn the optical response in a generalised way. We believe that these ideas are very applicable to many other problems in physics. We are currently working on further generalising the neural network's understanding of light-matter interaction with the long term goal to reach an ultra-fast model, able to deal with multi-material systems, arbitrary illumination conditions and possibly large-area geometries such as entire photonic meta-surfaces."
The new study included the latest developments in deep learning neural networks to deal with the multi-dimensional and large data-sets. The neural network was run on a graphics processing unit (GPU), awarded by industry partner NVIDIA to the project. The lead author was supported by an international fellowship from the German Research Foundation (DFG).