Faster and more energy-efficient with light

Artificial intelligence (AI) is increasingly shaping our everyday lives: chatbots answer service requests, AI formulates texts and designs images, and facial recognition unlocks millions of smartphones worldwide. However, the computers required for executing these complex tasks are becoming increasingly larger, expensive and energy-hungry. Hence, modern AI models are pushing conventional hardware to its limits—a problem that will become more severe in the future. 

To address this problem, the scientists of a new junior research group at Friedrich Schiller University Jena want to develop optical computers that work with light instead of electricity and use computing units as small as the atomic building blocks of crystalline materials. This concept could significantly increase AI computing power and open up an extremely energy-efficient, compact and sustainable alternative to today’s supercomputers. The Federal Ministry of Research, Technology and Space is funding the project »PicPhotMat —Structured Materials for Picophotonic Analogue Computing« as part of its »NanoMatFutur« programme with a total of 2.3 million euros over five years.

»Modern AI methods are based on neural networks and achieve brain-like performance in tasks that are challenging for traditional computers but easy for humans to accomplish,« explains Dr Elena Goi, head of the new junior research group at the University of Jena. »Optical systems are ideal for quickly processing the large amounts of data that neuromorphic—i.e. brain-inspired—systems deal with. They can perform complex mathematical operations on large matrices, which are the basic building blocks of AI algorithms, with significantly higher energy efficiency compared to conventional processors, outperforming silicon-based electronic technologies commonly used today.«

One million times smaller than the diameter of a human hair

Despite their great potential, these neuromorphic optical systems are not yet ready for broad applications. A main reason for this is that these systems are difficult to miniaturize and integrate on a computer chip in large numbers. Currently, artificial optical neurons still require a lot of space on a chip, even with state-of-the-art manufacturing technique.

Although so-called metamaterials allow for very compact elements for information processing, these are typically static and must be tailored for each application. This is not practical for versatile AI systems that need to respond flexibly to different tasks. This is where »PicPhotMat« comes in, exploring new ways to design miniaturized optical components that are as versatile as today’s electronic computer chips.

»Our project builds on the latest findings in pico-photonics, a field of research that investigates the interaction between light and matter at distances a million times smaller than the diameter of a human hair,« explains the physicist from the University of Jena. »The interaction of light and matter on this scale has the potential to influence the behaviour of light in powerful ways. This opens up completely new possibilities for the development of highly compact, energy-efficient and scalable optical neuromorphic systems.«

With this special technology, the »PicPhotMat« project aims to pave the way for sustainable and powerful AI hardware for the future. The researchers want to develop chips that require less energy while delivering significantly higher performance than today’s systems. In addition, the project aims to provide new insights into light-matter interaction at the smallest levels—knowledge that could open up completely new technical possibilities in the long term.