曾漢奇教授專訪 訪談：張芯瑀、陳宜榆 / 文：陳宜榆
Prof. Tsang’s research is mainly on photonic integrated circuits, particularly silicon photonics for high-capacity data center optical interconnects, and silicon photonics for optical sensing and imaging. His group has specialized in the technology of advanced waveguide grating couplers, setting records on optical bandwidths (>150nm) and lowest loss (<1dB per interface) using photolithographic patterning and processes available as standard PDK in commercial silicon photonic foundries. Current research projects include research on reconfigurable silicon photonic integrated circuits, hybrid integration of photonic materials with silicon photonics and integrated quantum photonic devices for quantum communications, quantum metrology and quantum computing.
You began your research in silicon photonics over two decades ago, long before it became a mainstream field. What drew you to it so early and how did you anticipate its potential?
A curiosity-driven research approach
To start, there are two ways of doing research. One is to look at the problems that people encounter currently and then to try to think of how to solve those problems. And that is applied research where you have a very directed objective of finding a new way to solve a particular problem. Then there is another type of research which is more curiosity driven. And the curiosity driven research is to ask a question about a certain material or a certain type of system to see if you can understand it better.
The way I got into silicon photonics was I started my research career working on indium gallium arsenide phosphide, semiconductor waveguides, not silicon. I studied those for maybe six, seven years. I did my Ph.D. in the University of Cambridge and after my Ph.D. I also worked for a while as a postdoc in the U.K. in the University of Bath and also visited Bell Labs in the U.S. where we tried to use the 3,5 semiconductor waveguides for making optical devices, such as tunable filters or for making optical switching using all optical switching. I discovered with 3,5 waveguides that there was a big problem in two photon absorption and free carrier absorption. That was the background to this. Another big problem I had was that it was very difficult to do the etching vertically. Every time you do the etching there's an angle, which makes it very difficult to control the device dimensions and the reflections on the device.
What I then encountered was some papers that suggested that silicon was a much more mature semiconductor technology where it was possible to etch waveguides perfectly vertically. And I thought: if it is so much easier to make perfect waveguides, perfectly etched and as designed, I asked the question whether it has the same problems as 3,5 waveguides had for two photon absorption and free carrier absorption. Around the year 2000, I started work on studying silicon waveguides and to try to address that problem. I did some basic experiments to look at the nonlinear effects in silicon waveguides and whether my hypothesis was correct.
In doing that research, there was a professor at UCLA who did a theoretical prediction that said that it was possible to make silicon optical amplifiers. Professor Bahram Jalali said that you can pump silicon waveguides and because it is a crystal, you can get high Raman gain and make an optical amplifier. He also published a paper on this, purely theoretical, and I looked at it and said, hold on, this doesn't seem right, because you have two photon absorption and free carrier absorption that will introduce nonlinear loss, in turn killing the gain. I then did some experiments with pulse pumping and then showed that this was indeed the case. This was one of my early papers that became very highly cited as we were the first group to say that silicon photonics has a high nonlinearity, but it also has high two photon absorption and high free carrier absorption that limits the amount of power you can send into the silicon waveguide. So that got me started and then we started doing work on trying to improve the silicon photonics, the coupling between chips and fiber.
The surge of significance of silicon photonics
Our group also started working on silicon modulators. The whole platform, although it's become much easier to manufacture the devices, there are still so many challenges in the technology that it just became interesting to try to solve some of them. And over time, I must have had probably close to 10, or more than 10 PhD students working on this topic from Raman amplifiers, gridding couplers, switching modulators, refractive index sensing, ring resonators, many of these devices that were developed. Because we were one of the first research groups in Asia and the world to work on these areas, it turns out that it was a very fruitful area for academic research. It turns out that now, after starting as a blue sky sort of curiosity driven research, 25 years later, there are some important real applications of silicon photonics.
It actually now underpins the way the internet functions. If you look at the traffic patterns of the internet, most of the traffic goes into a data center, maybe 90% or more. The traffic going into a data center volume is such that if the data center could not handle that amount of traffic in the internet protocol, (information) packets would need to be dropped and re-sent. The delay would be very similar to a denial-of-service attack on the internet servers, and everything would come to a crashing halt. It would completely fail if the data centers cannot handle the volume of traffic. But luckily, silicon photonics transceivers with the many data lanes per integrated transceiver could do that. With the advent of these transceivers supporting the data centers, the internet traffic can continue to grow.
Another emerging application that you read about these days of silicon photonics is in the so-called co-packaged optics, CPO. If you look at NVIDIA, they have a YouTube research, NVIDIA YouTube GTC. They talked about launching their co-packaged optics. So what they're doing is for the GPU-to-GPU interconnect, they're using silicon photonics to try to reduce the power consumption by a factor of three and a half and increase the data capacity to allow more data transmission and higher performance. This is now getting to the stage where in the future, you can imagine every single GPU in a cluster will have a co-packaged silicon photonics chip beside it to do the data transmission.This will become an extremely large volume underpinning the AI revolution using GPUs. So not only the internet, but also GPUs are doing this.
Recently in Hong Kong, my PhD students and I established a spinoff company in silicon photonics called Optik. It only started operation two months ago, but we were able to raise $7 million in funding for three years. So the $7 million funding is leveraging on the two-thirds funding from Hong Kong government, one-third funding from investors. What we're doing in our company is exactly what I said on data center transceivers, on co-packaged optics.
We've also been working on using silicon photonics as spectrometers to measure the optical spectrum of light on the chip. And spectrometers are widely used in, for example, optical currents tomography. I'm not sure if you've studied this, but in optical currents tomography, one of the ways of measuring a 3D image is to use a broadband light source like an LED.
And as you know, when you have a very broadband light source, the coherence length of the light is very, very short, less than a fraction of a millimeter. By using an LED with, say, 80 nanometers spectral width, you can resolve the depth of information down to a few micrometers in depth. So one of the major applications of OCT is to look at the eye or underneath the skin, for skin cancer detection. Our dream is to build an OCT instrument that is as small as a chip-based instrument on a mobile phone for scanning cancer underneath the skin.
But it is also true to say that the number of universities training designers of silica photonics is still rather limited. There actually exists quite a severe downfall in the current market by employers in search for talent of experienced designers in silica photonics. In my group, I have a group of about 10 PhD students, and most of them have found it really easy to find jobs doing silica photonics design. Our alumni have ended up working in Cisco in the U.S., in Intel in the U.S. doing silica photonics, in Lumentum in the U.S. working in silica photonics. And also, with this aforementioned startup company in Hong Kong, I recently grabbed six of my PhD students to form this startup. I gave them the shares in the company to try to get them to build the next unicorn in silica photonics, so hopefully they can succeed to make a really successful company with this technology.
So that's my dream, that the technology will be so successful that maybe like Qualcomm, or if you look at the, I just visited Mediatek. It's a fabulous design house that grew very big in Taiwan. And maybe in this company, Optik, we can also become next, like Mediatek in Taiwan, as a silica photonics specialist to do so many fabulous designs for silica chips.
I think it's made parallel.
Today, there are foundries that can manufacture silica photonics chips, but not so many design houses or designers that can do that. The more design houses there are, the more applications we can create solutions to solve with silica photonics chips. By making this happen, we can actually improve the technology for many things, not just data communications, but in imaging, 3D imaging, in industrial metrology, in LiDAR systems, in the fibrotic gyros, in even signal processing that we can use silica photonics for.
So I think that's a long answer to your question. In Taiwan, mostly we're still focused on integrated electronic chips right now, and silicon photonics is not yet that much of a topic. It's not yet mainstream. It's emerging. If you were a graduate in 1985, and somebody was to tell you that you should go into designing ICs, and then there was a company in the US that seven people started called Qualcomm. They offer you a job, and they say, oh, we'll give you 10% of Qualcomm shares. And if you imagine, that's the situation that it is like that now. So it's an opportunity if you were to take the plunge of joining a startup and take the risk of making the startup a success. They may only have 5% or 2% of shares, but if it becomes big, they can become extremely wealthy in 10, 15 years when the company goes IPO. I think it's a similar situation.
Applications in the industry
There are many opportunities that exist today. If you follow the news, Computex, NVIDIA announcements of this trying to form a consortium for co-packaged optics, and then Intel with another competing consortium for co-packaged optics. You see all this, and then you look at the technology and realize that neither of them actually solve the biggest problem in the interconnects.
The biggest problem is the bandwidth density. The electronic GPUs today probably need upwards of 20 terabits per second per millimeter of data transmission, and the way they're doing it is they have the edge of the chip, and they have these optical fibers coming out. If you look at the size of the chip, one of the constraints is how much data per millimeter length of edge that they have. The current solutions by Broadcom, for example, is UCIE. And if you look at the spec, it is 1.3 terabits per second per millimeter of edge. But that is still a factor of 10 less than what is needed. It’s apparent that there's still a sizable disconnect with the industry and the needs of the technology.
To conclude, there are still a lot of open challenges that perhaps we can solve with better designs of silicon photonics. Many of these existing problems, even with companies tackling them, aren’t immediately solved, and this leaves opportunities for clever PhD students, clever postdocs to come up with even better solutions than what the big companies are doing, which is how we're positioning this tiny startup.
To be a bit more specific, a lot of your early work centers around nonlinear properties. We've learned in class that for nonlinear properties to exist, you would need to have a very high intensity, or high energy density input.
It's an interesting point. I attended a conference last month in the U.S. called CLEO, Conference on Laser and Electro-Optics. In one of the sessions, one of the speakers was talking about single photon nonlinear optics. Single photon is actually in the context of quantum photonics. The idea there is if you have a very short optical pulse, and determine the duration of the pulse down to femtoseconds, in that duration, you will only have one photon. If you calculate the energy of that photon dissipated over femtosecond, the PAR (power-average rate), which is the energy per unit time, with the time unit being very short, femtosecond, the peak intensity is very, very high. And if you then put that single photon into a waveguide where you confine it by 200 nm, the PAR density, the intensity would actually go into very high nonlinear optics.
You can start seeing nonlinear effects even when you have one photon coming in if you have a very short pulse. The thing you need to understand is: What is a photon? A photon is not identified by a single wavelength, nor is it identified by just a tiny particle. A photon could have a spread of wavelengths. It could, even if you have a 20, 30, 40, 50 nanometer spread of wavelengths, it could still be grouped together to be one photon. So a photon doesn't have to be a pure monochromatic wave. That's a new thing for us. A photon could be just one particle like this.
The other big thing about seeing photonics integrated in the future, I think, will be in the quantum computing, or the quantum information processing side. If you understand a bit about quantum mechanics, there is wave-particle duality. Instead of a wave of light in the Maxwell's equations, you think of light as particles. Consequently, if you send a photon into a silicon waveguide, it can interact with the electrons in the silicon system to excite the electron to a high energy state. The electron from that high energy virtual state can relax back down to its original state and give back out the photon again. So that's what normally happens.
However, in quantum mechanics, it is also possible, with a non-zero probability, that an electron in the virtual state can return back to the ground state and give out two photons. Is it simulated? No, this is the spontaneous parametric down conversion (SDPC), which is a second-order effect. It is also possible that the electron can absorb two photons to go to a high energy state and then return back again, giving out two photons.
The interesting thing is that the two photons emitted do not necessarily have to be identical to the two photons which are destroyed by the absorption process. They could be at slightly different wavelengths. For example, if you have a continuous light wave of 1550 nm going into the silicon waveguide, it excites the electrons in the silicon to a high energy state, and the electrons return back to the low energy state, give out two photons. Instead of two photons at 1550 nm, you can have one at 1549 nm and one at 1551 nm, as long as the total energy is the same as the original energy is conserved. This means that you can generate a pair of photons from those two uncorrelated photons.
Another intriguing fact is that the pair of photons are emitted at the same time, so they're correlated in time. The other thing to understand is that they are entangled states. They are one superposition state. Even though they appear as two photons that can go to different destinations, if you make a measurement of one of the photons, you will affect the other photon. This is the quantum effect, the non-local effect that you make a measurement in one lab and affect the photon that is 10 kilometers away in another lab. This is useful in terms of the quantum information processing and communications, and we can even use this for algorithms to do quantum simulation, and it is very easy to observe in silicon waveguides.
So far there are a couple of companies now who've raised over 250 million or 800 million US dollars each called SciQuantum in Silicon Valley, Sanadu in Canada, that are trying to develop the silicon photonics for this quantum information processing. That may be a technology that will come five, ten years down the road, coming from silicon photonics. Silicon photonics is not just data communications in data centers, or GPU, or for spectrometers, or OCT, or for phased array for LiDAR. It can also do things that you cannot yet imagine in terms of quantum communications or quantum computing. This could be the basis of the next generation of quantum computing, which is also a very, very new area.
What we've learned in class is that transatlantic fibres, they used to be transatlantic. The main goal would be to minimize the loss because it needs to transmit at a long distance. But now in data centers, they're transmitting in short distances, but we need to maximize the data we can transmit at a given time. Our next question would be, with this change, what was different between these two types of optical transmission?
If you look at telecom, long distance communications, there's typically based in the 1550 nm wavelength band because of the lowest loss in silica fiber in the past. It's at that wavelength, whereas in data centers, it is typically based around 1310 nm, which is the wavelength of serial dispersion for the high-speed communications. We have an effect of optical dispersion. When you send in a short pulse or a short set of pulse data, say, 200 Gb/s, the time window would be about five picoseconds or something of that order. The spectral width, if you do the time-bandwidth product, would mean that if you have dispersion, you could have the beating between the short and long wavelength part of the pulse, which would cause destructive interference at some distances, leading to a notch in the transmission window, and ultimately cause errors. Dispersion is actually a problem if you have very high data rates and very long lengths of fiber. The dispersion means that it is important to work at the close to the serial dispersion wavelength in the fiber, which is why most data center transceivers are working around the 1310 nanometer.
There is a new technology emerging called hollow core fiber, where it does not use glass. It is actually a hollow core and an anti-resonant fiber. They have recently been reported to be even lower loss than glass fiber. In the future, there may be possibilities to work at whatever wavelengths can support the highest power in the fiber, meaning that we may not necessarily be working at 1550 nm.
That is also emerging technology. Okay, I'll just ask the next question, but I'm not really sure if you've already talked about it. So if you have, just feel free to say it. Okay, so I've gone on to your website, and it says that... It's probably out of date, sorry. I haven't updated it for months and years. I think it is updated. I'm not really sure if it's the one, but it says that your recent work features like a leading modulator data rate.
The importance of reducing insertion loss
We have published data that goes up to over 90 GHz now. What we've published is 330 gigabits per second in PAM-8 modulators. But unpublished work from my students have gone beyond that, pushing the state of the art every time. However, I think the emphasis on data rate probably misses the point. You might not have noticed the more unassuming work on the grating couplers. Grating couplers are devices to couple from a chip to a single fiber. It seems pretty routine. Everyone uses it for testing, and it doesn't look particularly interesting. You have a light in a waveguide, and you diffract light into the fiber. Pretty standard.
But if you understand a bit more about the systems, you understand that the loss in the fiber-to-chip coupling is actually what I would regard as the most critical importance for silicon phonetics technology. Insertion loss has many implications. Let me put it this way. In a transceiver, in a network, what is really required is to support many parallel data lanes. One example of a transceiver is with a multi-ribbon fiber. You have eight fibers coming in and eight fibers coming out in the pluggable transceiver, in the MPO connector. The idea there is that you have one laser. You split it to eight channels. In each channel, you modulate a different data lane. For example, if each data lane is at 200 gigabits per second, you have eight channels, eight fibers. With one laser, you can support 1.6 terabits per second, eight times 200 G.
The problem arises is that if you start having modulators with an insertion loss of 6 dB, and each fiber-to-chip interface fiber coupling loss of another 3 or 4 dB loss, then you will have around 10 dB loss per modulator. Basically, if you also have a 1 to 8 splitter, then you will find that even if you have 100 mW (milliwatt), divide by 8, you have 12 mW, and then you have 10+ dB loss in the transmitter chain. You're starting to get less than a milliwatt going out into the fiber. And then if your loss goes up to 2 or 3 dB each time, then you only have maybe half a milliwatt, which is insufficient to support data communications.
What happens is that you then need to increase the number of lasers you have in your transceiver from 1 to 2 or 4, depending on loss. And this leads to a problem that every laser is a potential single point of failure for the transceiver. If you have four lasers in the transceiver, it's a probability problem should any of them fail. That means that the reliability of the system becomes much worse, and you have failures in the field of the transceiver. It all comes back down to the problem of coupling loss and how to reduce insertion loss.
That's why I'm probably most proud of the results of my group getting less than 0.9 dB loss in a standard gridding coupler. It said that's sub-decibel. It is something that doesn't look very significant to a non-engineer. But if you understand the impact on the system, it is extremely important technology. I regard it as so important that I asked the university to patent it. We actually did, and then we had some foundries actually asking us to license it, meaning that it is a very useful patent. From a generic point of view, some would think, oh, it's just another gridding coupler, and the innovation on how to make a low-loss gridding coupler doesn't seem that important. But it's a core technology in silicon electronics.
On the other hand, the high-speed modulators is something that probably anyone with the access to the particular foundry can do, given the right design. The only thing that is tricky in the design is to decide on the length of the ring, the gap between the ring and the waveguide. Those are the things that we can control as the layout. For a micro-ring module, you have a radius of the ring, a gap between the ring and the waveguide. Depending on the calculated loss of the radius and the length, you just do the design, and it will work. If you can do a simulation to do that, designing a 300G or 400G modulator is feasible with the right tools.
But to innovate a gridding coupler without understanding the physics of a gridding coupler is pretty difficult. People have tried to do inverse designs of gridding couplers and couldn't achieve what we did. If you do inverse design, you're lucky to get about 1 or 2 dB, and you have to go down to maybe 20-nanometer, 30-nanometer feature size. We did a design with 170-nanometer feature size that is manufacturable in the foundry that can be manufactured at high yield with no loss. So I think that's the key. I think people might be so impressed by the high data rate that if you don't reduce the loss, you'll be… The loss is usually the limiting factor.
The power, the loss, the reliability, everything comes down to insertion loss. If you can minimize the insertion loss for your photonic integrated circuit, usually that ends up to have so many beneficial side effects since it's actually the most important criteria in the system. To focus on insertion loss is the first lesson I learned from industry. It doesn't sound important, loss. People might think, oh, I just have an extra 1 dB loss, just spend 20% power. But once you add up the whole system budget, the extra 1 or 2 dB loss here and there quickly adds up. This is an important lesson that took me a long time to realize. If loss is too much, then everything is futile, basically. You're throwing away so much energy.
How might these theoretical discoveries be related to applications, like in data centers or such things?
All the theoretical work today in the data centers, in co-packaged optics, in spectrometers are very application oriented. They have some very specific engineering applications in mind. There are also interesting work on integrated quantum photonics, generation of squeeze light, generation of entangled photons with cluster states of 2, 4 or more photons, which are at the physics level of experiments to try to see if we can make efficient circuits that can generate cluster states for quantum simulations, which are not yet ready for applications because we haven't solved some of the problems yet, such as in improving the visibility of the quantum interference. Those, if we can solve, are, I think, some interesting challenges moving forward.
It's also interesting to note that even in the quantum photonics world, loss is also very critical. If you generate a curve of entangled photons and you lose one by random scattering, they're no longer entangled. Loss is also an extremely important thing to manage in the integrated quantum photonics to try to minimize loss to build a high-fidelity quantum circuit. It's important for both nonlinear quantum photonics and the linear quantum photonics.