Northeastern Researcher Uses Photonic Metasurfaces to Slash AI Energy Demands

Artificial intelligence is hitting a physical wall. The electrical power required to run complex models and cool massive data centers creates unsustainable environmental and operational costs. To solve this, Professor Yongmin Liu of Northeastern University is developing a hardware alternative that swaps electricity for light.

Liu, who holds joint appointments in Mechanical and Industrial Engineering and Electrical and Computer Engineering, is leveraging photonic metasurfaces to build optical neural networks (ONNs). These metasurfaces consist of human-made arrays of nanostructures designed to precisely control the properties of light. By using photons to manipulate, detect, and process data rather than electrons, the system can perform machine learning and deep learning tasks with significantly higher speeds and lower energy consumption than current electronic platforms. "By using light, we can do machine learning and deep learning with extremely low power consumption," Liu stated.

The Electronic Bottleneck

Modern AI relies on electronic circuits where electrons move through silicon. This process generates substantial heat and requires immense amounts of electricity to maintain the throughput necessary for large-scale models. As AI scales, the energy required for data processing and the subsequent cooling of hardware have become primary constraints on the industry's growth. Photonics offers a way to bypass these limits by utilizing the inherent speed and efficiency of light, which can transmit data with far less resistance and heat generation than electrical currents.

Implications for AI Scaling

If successfully implemented, the shift from electronic to photonic platforms could fundamentally alter the trajectory of AI development. Reducing the carbon footprint and operational costs of data centers would allow for the scaling of more complex models without a proportional increase in energy demand. Furthermore, increasing processing speeds toward the speed of light would remove one of the most significant hardware bottlenecks currently hindering real-time AI applications and massive dataset processing.

Future Outlook

Supported by a grant from the National Science Foundation (NSF), Liu's research continues to refine how these nanostructures can be optimized for complex AI tasks. While the transition from traditional silicon-based electronics to optical neural networks represents a major architectural shift, the potential for a drastic reduction in power consumption makes it a critical area of study. The next phase of development will focus on the integration of these photonic metasurfaces into existing computing infrastructures to determine how effectively they can replace or augment current electronic AI accelerators.