I recently came across an article titled Generative Design Could Radically Transform the Look of Our World. No, the article wasn't about semiconductor design... at least not yet. It was about architecture.

And chairs. For example, the image below shows a chair designed using this approach. Starting from a crude chair, the algorithm was instructed to optimize both the weight of the chair and what I think is a measure of stiffness. but since it is in um like we used to measure semiconductor transistor dimensions maybe it should be microns per gram or something. Wrong units are a journalistic everyday event, see How Many Journalists per Square Acre? for more on that pet peeve). As JE Gordon pointed out in his marvelous book Structures: Or Why Things Don't Fall Down, furniture rarely breaks because it is not strong enough, it breaks because it is not rigid enough. That's one reason regular carpenters make poor wooden shipbuilders since ships really do sink because their joints are not strong enough (and perhaps too rigid).

Obviously, some of this technology is driven by what is called additive manufacturing, or often just 3D printing. There's little point in designing a chair like the one on the right if there is no way to manufacture it. Conversely, once you have additive manufacturing, there is no point in sticking to designs that were designed to be easy to make with injection molding (or even traditional carpentry).

The subtitle of the article about generative design was:

The future is lightweight, curvy, and full of holes.

That immediately made me think of silicon photonics, and Cadence's layout product in the space CurvyCore (see my post Yoga is Passé, the Future Is CurvyCore).

Inverse Design

It turns out that a similar approach seems to work for photonic designs, although it goes under the name inverse design: you say what you want the photonic device to do, and then the system experiments until it hones in on an approach that works. It also turns out that, like the chair example I pictured above, you end up with designs no human would come up with. They are literally superhuman designs.

A human designer of a photonic device proceeds a little as a traditional chair designer would. Start from a basic architecture that is mostly straight lines or simple curves. Vary things like the width of the device, a splitter, say. Then simulate. Adjust again.

Now watch this video. The thumbnail is the final design, and the video shows how it was created starting from a very basic layout. The secret is in all the little lumps back where the light is first split. As it was put to me, "it is going past the limitations of the human brain."

What you are seeing is inverse design creating a Y-splitter. Normally, a human would aim for a single major bandwidth and a limited temperature range. The final device, with seemingly random wiggly curves, is functional across all three major bandwidths, so is higher bandwidth. It is smaller and works over a wider range of temperature. It is not a small difference. This is a combination of a SKILL P-cell in Virtuoso and the same model, written in Python, running in Lumerical's photonic simulator FTTD simulation with 3D finite element analysis. The layout being used in both environments is exactly the same.

This second video shows the dashboard of the same iteration taking place.

As Gilles Lamant put it to me when I talked to him about this:

This isn't just a university project, although started at Berkeley. It is in production and we can actually put those layouts in as part of a design. It is getting a lot of attention (on both the optimization side and on our side).

Analog

Of course, an obvious question to ask is whether these techniques work for analog design. Automating analog design has been a hill that many companies have died on over the years, perhaps most visibly Barcelona Design, which raised nearly $50M and had Joe Costello as its high-profile chairman. But perhaps the time is right. Just as neural networks had to wait for massive amounts of compute power to be available through GPU-assisted datacenter servers, perhaps all that compute power will work for analog and mixed-signal design too.

As it happens, Cadence is working on a project looking at just this. See my post Cadence is MAGESTIC.  The partners with Cadence are NVIDIA (maker of GPUs) and Carnegie-Mellon University (CMU), one of the top handful of computer science departments in the country.

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.