Caltech Brings Fiber-Optic Performance to Silicon Chips: A 20x Leap for Visible-Light Photonics

Posted on 22nd Aug 2026 06:05:28 in Artificial Intelligence, Machine Learning

Tagged as: silicon photonics, photonic chips, optical fiber, AI data centers, quantum computing

Optical fiber carries nearly all of the world's data traffic, and it does so with an efficiency that still seems almost impossible: light sent through hundreds of kilometers of glass arrives at the other end with almost nothing lost along the way. That performance comes from two properties — the glass is extraordinarily pure, and its surface is engineered to be smooth at an atomic scale. For years, engineers have dreamed of bringing that same performance onto silicon chips, where light could be routed through microscopic pathways instead of long spools of cable. A team at Caltech has now done exactly that, and the result could reshape everything from AI data centers to quantum computers.

In a paper published in the journal Nature, researchers in Kerry Vahala's lab at the California Institute of Technology describe a method for building on-chip waveguides — the nanoscale pathways that channel light inside photonic chips — from germano-silicate, the very same glass used in optical fiber. The work, led by postdoctoral scholar Hao-Jing Chen and graduate student Kellan Colburn, demonstrates signal loss so low at visible wavelengths that it surpasses the previous record held by silicon nitride devices by a factor of 20.

The announcement has drawn wide attention across the photonics community, and for good reason. Photonic integrated circuits (PICs) that can carry light with fiber-like efficiency could unlock a new generation of ultra-coherent lasers, chip-scale atomic clocks and sensors, and dramatically more energy-efficient communication links inside data centers — including the enormous facilities that train and run today's AI models.

Why Loss Is the Enemy of On-Chip Light

To understand why this breakthrough matters, it helps to understand the problem of signal loss. When light travels through any material, a fraction of its energy is absorbed, scattered, or otherwise dissipated. In optical fiber, that loss is vanishingly small — on the order of a few percent per kilometer. In conventional on-chip waveguides, the story is very different. The surfaces of tiny silicon or silicon-nitride pathways are rough at the microscopic level, and every imperfection scatters a little light. Over the short distances of a chip the loss may seem small, but many photonic devices rely on light circulating thousands or even millions of times around tiny rings, where minuscule per-pass loss compounds rapidly.

"For years, we have been working to translate the spool-based fabrication of optical fiber onto silicon wafers, while trying to preserve the fiber's hallmark of ultralow loss," said Kerry Vahala, the Ted and Ginger Jenkins Professor of Information Science and Technology and Applied Physics at Caltech. "We have developed a method to print optical circuits, made from the same material as optical fiber, directly onto the same 8- and 12-inch wafers used for computer chips."

The payoff of lower loss is not linear — it is dramatic. In lasers built on these resonators, every factor of 10 reduction in loss translates into a factor of 100 improvement in coherence, the property that determines how precisely and stably a laser's light oscillates. That compounding effect is why researchers obsess over fractions of a percent of loss per meter.

From Fiber Spools to Spiral Waveguides

The Caltech team's approach starts with the material itself. Optical fiber is made of germano-silicate glass, chosen over decades for its purity and low absorption. The researchers adapted that same material to a lithography-based fabrication process — the standard technique used to build computer chips — so that waveguide circuits can be printed directly onto silicon wafers.

Just as important is the geometry. Instead of laying waveguides out in straight lines, the team arranges them in spirals. A spiral lets light travel a much longer optical path while staying inside a tiny footprint, in the same way that a fiber-optic cable is wound around a spool. But because the spiral is drawn with nanofabrication techniques, the entire path fits into a few square millimeters — or less — of silicon.

That packaging matters enormously for real-world systems. "Germano-silicate waveguides demonstrate extremely low loss and are also readily adaptable to efficiently transfer light between optical fibers and semiconductor lasers, which is of paramount importance in reducing the overall energy cost of server infrastructure," said Henry Blauvelt, chief technology officer at photonic-circuits specialist Emcore and a co-author of the paper.

The Furnace Reflow: Smoothing Glass to the Atomic Scale

Even with the right glass, surface roughness remains the enemy. Scattering from microscopic bumps and defects on a waveguide's surface has historically capped the performance of visible-light photonic chips. The Caltech team found a way around that limit by exploiting a quirk of germano-silicate: its comparatively low melting temperature.

"Due to the comparatively low melting temperature of the material, we can put our devices into a furnace to 'reflow' the surface of our waveguides to get their smoothness down to the level of individual atoms, which largely suppresses the severe scattering loss that has limited conventional visible PICs," explained Hao-Jing Chen. The reflow step essentially melts the surface of the glass just enough to let it settle into a perfectly smooth finish — a trick that would not work with the higher-melting materials used in older designs.

The result is a platform that matches the performance of the best silicon-nitride devices at near-infrared wavelengths — the region where most telecommunications happens today — and blows past them in the visible part of the spectrum. "At visible wavelengths, our recent platform exceeds silicon nitride's record by a factor of 20, and we have more room to improve," Chen added.

Why a 2-Centimeter Chip Needs Kilometer-Scale Performance

At first glance, chasing kilometer-scale loss specifications on a chip that measures only about two centimeters across sounds like over-engineering. Colburn, the paper's co-lead author, admits the goal can seem "a little ridiculous" — until you consider how photonic devices actually work.

The canonical example is the ring resonator, a basic building block of modern optics. Light enters a circular waveguide and keeps circulating, lap after lap, with each pass resonantly enhancing particular frequencies. The physical ring may be just a few millimeters wide, but the effective distance the light travels is set by how much energy leaks out on each lap. In a low-loss waveguide, light can circle tens of thousands of times, accumulating the equivalent of meters — or ultimately kilometers — of optical path inside a device smaller than a fingernail.

"That's where low loss over meters, or ultimately kilometers, really matters," Colburn said. "The longer light can circulate, the higher the performance of resulting devices can be." The team demonstrated the point by building several devices on the new platform, including ring resonators, multiple types of lasers, and nonlinear resonators that generate a broad range of frequencies.

From AI Data Centers to Quantum Computers

The applications span far beyond the laboratory. Vahala describes the platform as having "a Swiss Army-knife quality" because the same low-loss waveguides can be applied in radically different settings.

• AI data centers: Photonic links that couple optical fiber to on-chip lasers with minimal loss could cut the energy consumed by server-to-server communication, a growing share of data center power budgets as AI workloads scale.
• Chip-scale atomic sensors and optical clocks: The expanded wavelength coverage in the visible band supports the atomic transitions used in precision timing, gyroscopes for rotation sensing, and ion-trap systems — shrinking room-sized laboratory instruments toward chip scale.
• Quantum computing: Ultralow-loss visible-light circuits are a key ingredient for photonic approaches to quantum information processing, where preserving delicate quantum states depends on guiding photons without scattering.

Lasers built on the platform already show more than a 100-fold improvement in coherence time over earlier designs, a figure that speaks directly to the precision possible in sensing and metrology applications. And the researchers are candid that this is a milestone, not a destination. "We haven't gone as far as we want to go, but we've made significant progress over the last five years, and that's what we're reporting on here," Vahala said.

For the broader technology industry, the takeaway is simple: the gap between the world's best long-haul light transport and the microscopic circuits inside computers just got a lot smaller. If the platform scales the way the Caltech team expects, the next generation of photonic chips may finally deliver the efficiency of optical fiber in the footprint of a microchip.

Sources

• Caltech — Extending Optical Fiber's Ultralow Loss Performance to Photonic Chips
• Nature — Towards fibre-like loss for photonic integration from violet to near-infrared
• ScienceDaily — Caltech breakthrough brings fiber-optic performance to silicon chips
• SciTechDaily — Caltech Breakthrough Brings Fiber-Optic Performance to Silicon Chips
• Photonics Spectra — Caltech Team Charts Path to Ultra-Efficient PICs