Source: SemiInsights  23^rd Sep 2026

For a long time, silicon photonics has integrated nearly every core component needed for optical communications—waveguides, modulators, beam splitters, and detectors, yet it has always lacked one critical element: the optical isolator.

Its function seems simple, but it is irreplaceable.

Semiconductor lasers are the light source in optical communication systems, but they are also highly susceptible to reflected light. Even extremely weak reflections from fiber end faces, connectors, or waveguide interfaces can disturb the laser’s oscillation state, causing wavelength drift and increased phase noise, and in severe cases even triggering communication bit errors. The optical isolator is responsible for blocking this backward light and allowing light to propagate in only one direction, which is why almost all high-performance optical communication systems depend on it.

Yet this seemingly ordinary device has become one of the hardest challenges to overcome in monolithic silicon photonics integration over the past two decades. 

Recently, a team led by Associate Professor Taichi Goto of the Research Institute of Electrical Communication at Tohoku University, together with Kyocera, announced that they have developed a new process for fabricating on-chip optical isolators using localized laser annealing, and successfully completed a monolithic silicon photonics integration prototype. The results were published in IEEE Access, and Kyocera described the work as the world’s first laser-annealing-based integration technology for silicon photonics optical isolators.

The real breakthrough is not just a device, but a manufacturing method.

Why Has the Optical Isolator Remained So Difficult to Integrate?

The problem lies in the material.

The best-performing optical isolators today generally use cerium-doped yttrium iron garnet (Ce:YIG) as the magneto-optical material. It can produce the Faraday rotation effect, causing forward- and backward-propagating light to exhibit different polarization changes, thereby enabling one-way transmission of light.

But Ce:YIG has a fatal flaw: immediately after deposition it is completely amorphous and has no magneto-optical properties. It must undergo high-temperature annealing at roughly 600–800°C to form a garnet crystal.

This creates a natural conflict with silicon photonics manufacturing.

Today’s data center silicon photonic chips, especially those using co-packaged optics (CPO) architectures, already contain large amounts of copper and aluminum multilayer metal wiring, electrodes, and dielectric structures. Above about 400°C, these metals may diffuse and deform, affecting conductivity; the interfaces between silicon waveguides and silica cladding can also become rough at high temperatures, increasing propagation loss.

In other words, as long as conventional furnaces are used to anneal the entire wafer, optical isolators and advanced silicon photonic circuits cannot coexist.

For years, therefore, the industry has had only one alternative: fabricate the optical isolator separately and then bond it to the silicon photonic chip through hybrid integration. This is why the optical isolator has long been called the “last threshold” of silicon photonics.

Using a Laser to Heat only Tens of Microns

The approach proposed by Tohoku University and Kyocera is straightforward: if the entire chip cannot be heated, then heat only the optical isolator itself.

The researchers first etched a microgroove into a commercial SOI silicon photonic chip, then directly deposited a Ce:YIG thin film by high-frequency ion beam sputtering, without a conventional seed layer.

They then used a semiconductor laser with a central wavelength of 915 nm and a maximum output of 120 W to precisely irradiate the groove area with a uniform spot about 700 μm square, bringing the Ce:YIG locally to the temperature required for crystallization while keeping the rest of the chip at a lower temperature. The entire process was carried out in a vacuum, with the substrate maintained at only about 250°C, greatly reducing the thermal impact on metal interconnects and silicon waveguides.

Compared with conventional millimeter-scale laser annealing, this micron-scale localized irradiation also solves another problem: film cracking caused by thermal expansion mismatch. Because the heated area is confined to a groove only tens to hundreds of microns across, stress between Ce:YIG and silicon is effectively released, and no obvious cracks were observed in the experiments.

This is the real innovation of the work: not improving the Ce:YIG material itself, but redesigning how it is made compatible with silicon photonics manufacturing.

AMZI: Allowing the Isolation Effect to Be Amplified

The experimental device uses an asymmetric Mach–Zehnder interferometer (AMZI) structure.

The silicon waveguide is 550 nm wide and 180 nm high, with a deliberately introduced optical path difference of 15 μm between the two interference arms. Ce:YIG grooves with lengths of 50, 100, and 200 μm are fabricated on one of the arms.

The working principle is not to directly measure Faraday rotation. Instead, a magnetic field is used to change the polarization state in the Ce:YIG, causing a wavelength shift in the interference spectrum, from which the magneto-optical rotation capability is then derived.

To avoid errors caused by re-inserting the fiber, the research team kept the chip and fiber positions completely fixed and simply reversed the polarity of a SmCo permanent magnet to obtain forward and reverse transmission spectra separately.

This measurement method offers higher consistency than traditional insertion-and-removal measurements and is also more representative of how an on-chip device actually operates.

What Performance Level Was Achieved?

The best overall performance was obtained with the 100 μm groove sample.

At the 1540 nm communication wavelength, optical isolation reached 13.6 dB, meaning the backward reflected light power was reduced by approximately 95%.

Even more noteworthy is the magneto-optical performance itself.

The researchers observed a 4.5 nm shift in the transmission spectrum peak. Combined with a free spectral range (FSR) of 44 nm, this gives a Faraday rotation capability of 0.092°/μm for the Ce:YIG, roughly nine times that of previously laser-annealed Ce:YIG and approaching the level of magneto-optical materials prepared by conventional high-temperature furnace annealing.

Transmission electron microscopy (TEM) further confirmed that the material had successfully crystallized. The 195 nm-thick Ce:YIG film formed a complete garnet crystal, with a lattice spacing of 0.273 nm for its (420) plane, highly consistent with the theoretical bulk value of 0.278 nm. EDX elemental analysis also showed a Ce:Y:Fe ratio of approximately 1:2:5, broadly consistent with the garnet stoichiometry.

These results show that localized laser annealing can not only achieve crystallization but also bring material quality close to that of conventional processes.

The Biggest Problem Is Not Isolation After All

What truly limits commercialization is insertion loss.

In the experiment, the device's forward insertion loss was as high as 20.4 dB, meaning most of the signal had already been attenuated after passing through the isolator. This is far above what data center optical interconnects can accept.

The paper further breaks down the sources of loss: about 66% comes from sidewall and interface roughness caused by groove fabrication, about 27% from absorption and scattering in the Ce:YIG itself, and only about 7% from the AMZI interference structure.

In other words, the biggest bottleneck at present is not the magneto-optical material but the precision of the manufacturing process.

The research team proposed several directions for further optimization, including improving the dry etching selectivity, reducing over-etching of the silicon waveguide, and controlling the oxygen partial pressure during annealing to optimize the cerium ion valence state. However, these proposals remain at the discussion stage and have not yet been experimentally validated.

Two More Steps Remain Before a True On-Chip Optical Isolator

Beyond loss, magnetic biasing remains another key challenge.

This experiment relied on an external SmCo permanent magnet to provide the saturation magnetic field. A true on-chip optical isolator deployable in CPO would also require the magnetic bias layer to be monolithically integrated directly into the chip. Otherwise, an external magnet would still be needed, making it difficult to realize the advantages of highly integrated packaging.