Artificial intelligence has rapidly transformed modern life, but behind its brilliance lies a pressing challenge: enormous energy consumption. Today’s AI models demand billions of parameters and vast computing power, pushing the limits of existing hardware and consuming energy rivaling that of entire cities. In response, scientists are seeking smarter, more sustainable approaches—and light may be the answer.
Researchers from the Swiss Federal Institute of Technology Lausanne (EPFL) have introduced a breakthrough in optical neural networks (ONNs), proposing a new computing model that replaces electrons with photons. Their method leverages light pulses traveling through multimode optical fibers, enabling powerful computations using far less energy than traditional systems.
Harnessing Light for Smarter Machines
The team, led by Professors Demetri Psaltis and Christophe Moser, developed an architecture that uses wavefront shaping to precisely guide ultrashort pulses of light through optical fibers. This allows the system to perform nonlinear computations, which are essential for modern AI tasks such as image classification and pattern recognition.
Unlike conventional AI hardware that requires massive data centers and consumes large amounts of memory, the optical system relies on a minimal set of programmable parameters. Remarkably, it achieves performance levels comparable to fully digital networks that rely on over a hundred times more adjustable weights.
Nature as the Hardware
Rather than building complex chips, the researchers found a way to utilize natural optical behaviors as the foundation for computation. “We discovered that we could select a set of weights—essentially the brain of the AI—from a natural ‘weight bank’ created by the fiber’s physical properties,” explained Ilker Oguz, lead author of the study. This eliminates the need for dedicated fabrication or energy-hungry digital circuits.
Because the system operates with microwatts of optical power, it introduces a viable route to building low-energy, scalable AI processors. More impressively, the team showed that their approach could be generalized to other high-dimensional, nonlinear phenomena—opening the door to diverse applications across scientific fields.
The Road to Green AI
As machine learning models grow in size and complexity, the need for efficient computation becomes increasingly urgent. Optical neural networks, like the one demonstrated in this study, provide a compelling alternative. By performing calculations through the natural movement of light, they drastically reduce the demand for memory, electricity, and physical hardware.
This research represents a crucial shift toward sustainable AI development—one that doesn’t sacrifice performance for efficiency. With further refinement, such systems could lead to compact, eco-friendly AI devices capable of operating on the edge, far from massive cloud servers.
Conclusion
The future of artificial intelligence may not lie in bigger chips or faster processors, but in light itself. Through the pioneering work of optical computing and intelligent wavefront control, we’re moving closer to AI systems that are not only smart but also sustainable. This marks a significant step in creating technology that aligns with both our ambitions and the planet’s needs.
