Automated design process creates smallest, most efficient photonic switch ever

Automated design process creates smallest, most efficient photonic switch ever
By using an automated design algorithm, researchers at Stanford University have built the world’s smallest and most efficient silicon based photonic wavelength splitter, or demultiplexer.

By using an automated design algorithm, researchers at Stanford University have built the world’s smallest and most efficient silicon based photonic wavelength splitter, or demultiplexer. If you haven’t come across that term before, for computers to communicate with each other, multiplexing and demultiplexing is essential. Multiple information streams are combined ( multiplexed ), transmitted over some distance, and then decomposed into their original forms ( demultiplexed ). Thanks to this concept, we can send extreme amounts of data over the Internet using optical fibers.

Let’s say we want to send 10 pieces of information over the Internet. To do so, we can use 10 different wavelengths of light to encode this information, send it down a fiber, and then reverse the process on the other end to decompose which wavelength corresponds to which information stream. (This technique is just one type of multiplexing ( wavelength division multiplexing) — there are many more.)

Now, as communications get faster, photonic devices continue to shrink and become increasingly integrated with electronic components, hence the field of nanophotonics. Yet it becomes increasingly difficult to build devices for telecommunications — not just because of fabrication problems, but because light acts remarkably different at smaller scales.

So rather than having to design every component of a photonic circuit from scratch, a team at Stanford University has shown that by developing a clever optimization algorithm, it’s possible to create a highly efficient wavelength demultiplexer without any top-down design whatsoever. The device, based on silicon, can be used to demultiplex two wavelengths, 1.3 and 1.55 microns (which are standard in telecommunications). The design is also the smallest we’ve seen of this type, at only 2.8 square microns, and it exhibits the lowest insertion and crosstalk loss yet.

SEM image of the demultiplexer from Stanford

With this concept, light from the input waveguide enters the demultiplexer, which splits light (according to wavelength) into two waveguide output channels. Now, this may sound unremarkable, but when you actually see the structure which achieves this, it almost appears counter-intuitive that it’s the optimum structure. The structure is shown to the right under electron microscopy techniques.

What makes this research different is that the demultiplexer is based on a non-conventional, inverse design method . Generally, designing anything from mechanical bridges to electronic circuits relies on a top-down theoretical approach. That means certain equations and design principles are employed, and we keep having to adjust the structure by hand such that it’s able to perform adequately.

For example, a typical mechanical bridge may not the optimum way to get across a valley, even though it’s one possible solution. There is an alternative train of thought: use optimization algorithms . Let’s say you’re designing a device to solve a problem. You specify what functionality it should have, and in response, the algorithm searches over the entire parameter space, continually modifying the structure at hand to produce the desired outcome.

Getting back to the Stanford team, the researchers wanted to create an arbitrary structurethat can physically split input light into two separate wavelengths and send them down two separate waveguides. So they created an optimization algorithm, which then runs (using a first guess at the structure to start things off), constantly modifying the structure as it goes, recording the result, and determining whether the latest structure produces the optimum output. Eventually, it arrives at the optimal design. This type of approach (feedback loop), can be thought of as an optimization algorithm. It tries to find the best (optimized) solution to some problem where a rigorous approach is either not possible, computationally unfeasible, or even just produces non-workable solutions.

Now, from an electromagnetic point of view, the propagation of light will vary depending on the refractive index of the medium. It’s possible to use an optimization algorithm that constantly iterates the spatial refractive index profile of the volume, in order to ‘force’ the light to propagate in different directions, depending on its wavelength (in other words, splitting the light). The optical power of the two waveguide outputs can be monitored (at the desired wavelengths). Then, after many millions of iterations, an optimized design can be found. The process is summarized in the figure below.

Design process for demultiplexer

The design process for deriving this device is interesting. Implementing it in the real world is another matter, and what remains to be solved is twofold: how long does it take, and what algorithm to use? Firstly, with any design method based on the above principles, one quickly runs into the problem of practicality. To actually run the algorithm so that it concludes with a reasonable outcome, it takes a long time. With this example, the researchers said it took 36 hours to derive the optimized structure.

Final design

Secondly, is the outcome actually the ‘optimum’ solution? Simply put, the answer is maybe. There are a plethora of inverse design algorithms, from the method used here to using genetic algorithms, and it’s important that the solution found is not a local optimum, but a global optimum. Imagine you are deliveryman and need to visit a series of cities within a certain time period. There are many routes between them, and many ways of visiting all cities. So it’s possible to visit all of them using one particular path, but that path may not be the best solution. There may be one in which your journey is significantly reduced in distance (the optimum) — in other words, the well-known traveling salesman problem.

Nevertheless, there is vast potential for this automated design process. The researchers plan to utilize this technique to create faster and more efficient integrated photonic circuits in the future.