Using AI to tune quantum computers

3:18:21>> The negatives, there is one, and that's where my job comes in. >> Aha. >> Okay. If you've made it this long, you now will know what Christian >> Finally, you'll know what I do with my life when I'm not um in this podcast studio. [laughter] So, there's a skeleton in the closet here, >> okay? >> Which is I'm fabricating stuff at the limit of um fabrication. There's the these gates, these wires are tens of nanometers. They have to be done with this isotopically rich silicon. Um, and when I fabricate, you know, at the nanometer scale for transistors, if a single transistor has a few extra atoms, not a big deal.

3:19:02>> No big deal. >> Not a big [clears throat] deal. Here it is a big deal because if a single gate has a few extra atoms, is slightly bigger, slightly smaller. If there's a random charge noise hanging around, then the same settings that I used for one set, right? Let's say I plus plus.4 minus.5 plus.652 652, whatever. >> I can't use those same settings. >> Yeah. >> Everywhere in my device because every single one of these electrodes because of the fab disorder, there's going to be inherent noise. >> Yeah. >> In how that happens, right? So, you would have to tune up this device,

3:19:42right? You'd have to for every single cubit, you'd have to figure out exactly the right settings to get three isolated [clears throat] >> to to understand how long I have to wait to to make an exchange gate to understand um what is the voltage what is the current when there are two when I do that like um poly spin blockade when I do the readout where I'm like shoving two electrons into a single thing how do I actually do that how long do I have to wait what corresponds to tworon electrons being in there versus one because I'm reading a current, right? >> And so >> you I mean you could have a PhD physicist do this by hand which fine for

3:20:22like small enough for the for the 2023 paper where there's like six electrons fine you could do that >> pretty soon if you want to scale this thing it is going to get insane right it would take a human lifetime to to tune up a million cubits by hand. Yeah. Right. >> Yeah. And so, so, and so you are the cubit tuner. >> Yes, I am part of the cubit tuning team. I I used to be part of the cubit tuning team at HRL. And that's still something that I do at Durac now. >> QT. >> So, so here's here's how it actually works. I'll just show you some of the supplement of this paper that goes into that kind of stuff. The first thing you do is you make a a dot charge sensor is what they call it.

3:21:03This is your single electron transistor. So you initiate the single electron transistor. That way you can sense what the hell is happening in my device. Then I start loading my dots. Before there's no electrons and then now I want to load exactly one electron into each or in some cases an odd number of electrons into each. That way the electrons pair up and only the thing that is like missing is going to be talking and doing the stuff. In order to do that what I do is I toggle the voltages. And when I toggle the voltages, the electrons are going to tunnel and go ploop into one, ploop into another. And every time that happens, my single electron transistor is going to move. And you see on the bottom the the C and D.

3:21:45>> Yes. >> Those are images that show these tunneling events. >> Like every time an electron is moving from one room to another, >> the the charge sensor is going to move. And so my current is going to be a different color. uh it's going to be a different value that's going to manifest as a different color. And so the two axes are my voltages. Like I'm scanning this voltage and I'm scanning this voltage and there's going to be an event that's going to show up as a streak. >> And those streaks are going to tell me what the settings are that I need to implement. >> Okay. Finally, once I have the correct settings, then I go into creating my poly splin ball blockade, my uh readout where I shove two electrons into a single confinement well to see if

3:22:26there's two electrons, then I get one value of current. That's the histogram on the right hand side. If there's one electron, then I get another value of current. And so those two then tell me, hey, if I got this current, that means I'm a one. If I got this current, that means I'm a zero. And then finally I then tune the other axis to get full block sphere control. So this is the sort of path that you would have to take for every single cubit >> in order to make it work >> right and each of them are have their own subsequent substeps like that's a very brief >> that's a very brief but this whole thing is going to take like two hours right for like a human being. >> Okay. >> Okay. Ain't nobody got time for that. >> Right. Right. Right. Right. For each for each cubit. >> Yeah. For each cubit. Right.

3:23:07>> Um >> so you use AI. >> Yeah. Of course. Of course. >> Of course. >> Of course. >> Um especially because like the the bottom part where you're like finding the edges and you're finding the like >> it's just an image, right? >> That I'm looking for streaks, >> right? >> Object detection. >> Yeah. The the computers have vision now. >> Yeah. Yes. Exactly. Computer. So that that's that's that's kind of the some of the stuff that I worked on with the automation team. >> Makes sense. um is you know this is uh this is the machine learning sort of road map on how to get from the scan >> where you get the these tunneling events all the way to okay what are the voltages that create those tunneling events and then what are the settings that I need to create this is a um der

3:23:50>> um it's a detection transformer >> so it's a it's a hybrid between a convolutional neural like a resnet that actually does most of the object detection and stuff, but then you feed that stuff into a transformer then uses attention to then create like relationships between all of these lines to tell me where the stuff is. >> God, this is so good. >> Okay, so now you actually finally have have a good idea. >> This this chart I understand [laughter] this one this one makes sense. But I think it's it's a very what's really cool about or important about not important what's interesting about this is that of everything we talked about to

3:24:31get to this point every aspect of it is this complex >> and we're just talking about a a sliver a slice. >> That's that's a very good point. Yeah. every single aspect which is why there's 250 authors >> because everything is so like you have to be so deeply understanding of the the physics aspects of it and then whatever the applied avenue that you're dealing with it are you materials person are you uh working in algorithms and computation at a chip level are you working on those at the the external are you are you building the software that lets me right like do this >> do do this actually because it's just because you have the chip doesn't mean you with it. There's so many and every one

3:25:12of them has a depth of knowledge and execution necessary that I think people really can underestimate when we don't work in these things and we sort of will go through a paper and we'll explain it and it's like yeah we're taking away the things but like >> a lot of people have worked very very very very hard in an area that very few people can even work in. >> Yeah. >> Um to accomplish that. Uh so that's like because the tuning piece again it >> it's one sliver but it's a very important like it's like >> I mean it's part of the scalability argument. >> Right. Right. >> Right. Of the three layers of the criterion. Anyway, I'm just trying to give you props but like that >> it's dude this is this is so guys >> I'm I'm really actually just like on a

3:25:53personal note I'm very happy that we did this because now my podcast partner actually knows what I do on a date day as a day job. >> Right. And like now I know when we go out and we're people I'm going be like you know what he does >> and I'll be able to give him the spiel. >> Yeah. Exactly. [laughter] So that was the control part. And although towards the end we kind of got into scalability which is like if we want to build a bigger thing you better be able to tune that thing autonomously, right? You don't want a human being there sitting there like a monkey doing this. Um and so finally we got into scalability in economics. And this is the kill shot >> for silicon. Um, if you want to build a