What if I told you light can power your phone? I am not referring to your phone being charged via solar panels. I mean instead of electricity coursing through your phone’s circuits, it is light.

We are already living in a partially light or optics-powered society. Solar panels are rippling through rooftops in Southern California and during the summer, we can turn on the air conditioner without worrying about our electricity bills skyrocketing. Watching my favorite shows on Netflix has never been easier because I can easily VPN to Korea to watch my favorite Korean dramas. This is only possible because optical fiber cables transmit at the speed of light to bring streaming to us from continents away. Electricity, however, acts a little differently than light.

By now, we have probably all experienced our computer running too hot from too many programs open or our phone heating in our hands as we play games. This is because electricity naturally generates heat. As AI companies, like Nvidia and AMD, continue the race to make AI chips, their hardware is going to generate more heat. Instead of using electricity as its driving force, companies should look to using light because it will operate faster and minimize heat generation than its electronic counterpart.

This past spring, I had the privilege of studying nanophotonics at Georgia Tech. In this short article, I am going to explore how AI can drive advancements in optics and, conversely, how optical devices can enhance AI technology. Optical technologies can significantly benefit AI development by enabling faster and more energy efficient data processing and calculations. To achieve this goal, AI must first help propel optics innovation in the inverse design of devices with specific optical properties or by using surrogate modeling to predict light interactions with these devices.

Moore’s law is the observation that the number of transistors in a circuit will double every two years. This is the reason why computer chips are faster, but also the reason why their speeds are beginning to plateau. As more companies besides Nvidia develop chips for AI, how can companies stay ahead of its competition? How can we break through this barrier so that AI hardware can continue to grow? The answer is an optical neural network.

While this technology might be new to industry, researchers have been developing optical neural networks for years. In 2022, Luo et al., creates an optical neural network that incorporates a complementary metal-oxide semiconductor (CMOS) to sense light. In combination with the use of AI designed surfaces, this optical neural network is capable of parallel processing and is easily scalable for real world applications (Luo 2022). Another instance of optical neural network is created by Qian et al. where they design a diffractive neural network that can utilize all 7 logical operators. AI designed surfaces can redirect light propagation onto designated 1's and 0's regions of the output layer to mimic the 7 different logical operators (Qian 2020).

Before we can create an industry ready optical neural network, we need to first better understand optical devices via inverse design and surrogate modeling. An example I want to showcase is the design of an invisible cloak, like the one in Harry Potter. Although invisible cloaks are not new technology, the example below follows one created with AI assistance.

The experiment by Zhen et al. in 2021 utilizes a tandem neural network (TNN) comprise of a forward network for surrogate modeling and a reverse network for inverse design. The TNN's forward network predicts light interactions with a surface, then feeds this information back into the reverse network to iteratively refine the surface design. This process continues in a loop until the desired outcome is achieved—in this case, the optical concealment of a cat. The experiment employs two AI-designed surfaces to achieve this concealment, showcasing AI's potential in enhancing our understanding of light-matter interactions (Zhen 2021). With a deeper understanding of light-matter interactions, we can create more intricate devices that can slowly replace electronic components.

While a complete transition to an optical-based society may not happen within our lifetime, we will undoubtedly see a rise in optical and electronic hybrid devices. As we navigate the age of AI, it is crucial for AI hardware to evolve and meet increasing demands. In a world shaped by vast opportunities and high expectations, it is essential to innovate our foundations so AI can continue to positively impact other industries.

Zheng, Zhen, et al. "Realizing Transmitted Metasurface Cloak by a Tandem Neural Network." Photonics Research, vol. 9, 2021, pp. B229-B235.

Rowling, J. K. "How Harry Potter Got His Invisibility Cloak." The Telegraph, www.telegraph.co.uk/culture/harry-potter/11882434/JK-Rowling-explains-how-Harry-Potter-got-his-Invisibility-Cloak.html.

Luo, X., et al. "Metasurface-Enabled On-Chip Multiplexed Diffractive Neural Networks in the Visible." Light: Science & Applications, vol. 11, 2022, p. 158, https://doi.org/10.1038/s41377-022-00844-2.

Qian, C., et al. "Performing Optical Logic Operations by a Diffractive Neural Network." Light: Science & Applications, vol. 9, 2020, p. 59, https://doi.org/10.1038/s41377-020-0303-2.

"How to Review 4 Million Lines of ATLAS Code." Scientific Figure on ResearchGate, www.researchgate.net/figure/CPU-transistor-densities-clock-speeds-power-and-performance-from-1970-2015-Courtesy-of_fig1_321233071. Accessed 6 June 2024.