Quantum computing and AI join forces for particle physics

1
00:00:08,000 --> 00:00:10,960
Hello, and welcome to the Physics World weekly

2
00:00:10,960 --> 00:00:12,900
podcast. I'm Hamish Johnston.

3
00:00:13,724 --> 00:00:14,544
This episode

4
00:00:14,925 --> 00:00:16,864
explores how quantum computing

5
00:00:17,244 --> 00:00:18,785
and artificial intelligence

6
00:00:19,324 --> 00:00:20,304
can be combined

7
00:00:20,684 --> 00:00:21,664
to help physicists

8
00:00:22,204 --> 00:00:24,304
search for rare interactions

9
00:00:24,925 --> 00:00:26,864
in data from an upgraded

10
00:00:27,244 --> 00:00:28,945
Large Hadron Collider.

11
00:00:29,989 --> 00:00:32,810
But first, I'd like to thank Delft Circuits

12
00:00:33,190 --> 00:00:35,929
for their generous support of this episode.

13
00:00:42,804 --> 00:00:45,545
This episode is supported by Delft circuits,

14
00:00:46,004 --> 00:00:48,184
a key enabler in the evolution

15
00:00:48,564 --> 00:00:49,945
of quantum technologies.

16
00:00:50,964 --> 00:00:54,824
While annealing quantum computers are already demonstrating

17
00:00:55,204 --> 00:00:56,265
practical value,

18
00:00:57,070 --> 00:00:59,869
Gate based systems are still on their way

19
00:00:59,869 --> 00:01:01,729
to large scale implementations.

20
00:01:02,909 --> 00:01:04,530
One of the biggest challenges

21
00:01:04,829 --> 00:01:05,810
is IO,

22
00:01:06,590 --> 00:01:09,329
reliably connecting thousands of qubits

23
00:01:09,734 --> 00:01:10,555
without compromising

24
00:01:11,015 --> 00:01:11,515
performance.

25
00:01:12,694 --> 00:01:13,674
Delft circuits

26
00:01:14,135 --> 00:01:16,715
addresses this bottleneck with a fundamentally

27
00:01:17,174 --> 00:01:19,275
different approach to cryogenic

28
00:01:19,975 --> 00:01:20,475
connectivity.

29
00:01:21,415 --> 00:01:22,234
Their flexible,

30
00:01:22,935 --> 00:01:23,435
multichannel

31
00:01:24,180 --> 00:01:25,240
planar cabling

32
00:01:25,859 --> 00:01:28,600
integrates filtering components and delivers

33
00:01:28,980 --> 00:01:29,480
exceptionally

34
00:01:29,859 --> 00:01:31,319
low heat load,

35
00:01:31,700 --> 00:01:34,760
paving the way to kilo cubit architectures

36
00:01:35,859 --> 00:01:37,640
because advancing science

37
00:01:37,939 --> 00:01:38,439
requires

38
00:01:39,204 --> 00:01:39,704
infrastructure

39
00:01:40,084 --> 00:01:42,185
built for the quantum age.

40
00:01:49,765 --> 00:01:53,340
In mid twenty twenty six, the Large Hadron

41
00:01:53,400 --> 00:01:56,859
Collider at CERN will shut down and undergo

42
00:01:57,079 --> 00:02:00,219
the final stage of the high luminosity

43
00:02:00,760 --> 00:02:01,260
upgrade.

44
00:02:01,959 --> 00:02:04,920
The aim is to increase the number of

45
00:02:04,920 --> 00:02:06,060
particle collisions

46
00:02:06,439 --> 00:02:07,340
in the LHC

47
00:02:08,055 --> 00:02:10,314
by about a factor of 10.

48
00:02:11,254 --> 00:02:12,635
By 2030,

49
00:02:12,854 --> 00:02:14,314
this should allow physicists

50
00:02:14,935 --> 00:02:16,715
to study rare particle

51
00:02:17,014 --> 00:02:17,514
interactions

52
00:02:17,974 --> 00:02:19,915
that are not visible today.

53
00:02:21,439 --> 00:02:22,340
These interactions

54
00:02:22,719 --> 00:02:24,580
create showers of particles

55
00:02:24,959 --> 00:02:25,939
that are detected

56
00:02:26,240 --> 00:02:27,780
by huge LHC

57
00:02:28,159 --> 00:02:28,659
experiments

58
00:02:29,280 --> 00:02:29,780
including

59
00:02:30,080 --> 00:02:31,539
ATLAS and CMS.

60
00:02:32,974 --> 00:02:34,914
These showers create distinctive

61
00:02:35,215 --> 00:02:35,715
patterns

62
00:02:36,014 --> 00:02:37,634
in the detector data

63
00:02:38,175 --> 00:02:39,555
which provide details

64
00:02:40,014 --> 00:02:42,834
of the sought after particle interactions.

65
00:02:44,334 --> 00:02:45,875
The challenge for physicists

66
00:02:46,449 --> 00:02:49,110
is to isolate the desired signals

67
00:02:49,489 --> 00:02:52,209
within the plethora of data that will be

68
00:02:52,209 --> 00:02:54,069
produced by an updated

69
00:02:54,370 --> 00:02:54,870
LHC.

70
00:02:56,129 --> 00:02:58,709
To do this, they must know how showers

71
00:02:58,769 --> 00:03:00,069
from rare interactions

72
00:03:00,689 --> 00:03:01,590
will interact

73
00:03:01,915 --> 00:03:02,735
with the detectors,

74
00:03:03,594 --> 00:03:05,294
and this is done using

75
00:03:05,675 --> 00:03:06,175
computationally

76
00:03:06,715 --> 00:03:07,215
intensive

77
00:03:08,075 --> 00:03:08,575
simulations.

78
00:03:10,075 --> 00:03:10,575
Therefore,

79
00:03:10,955 --> 00:03:13,615
it's not surprising that particle physicists

80
00:03:14,075 --> 00:03:16,094
are looking to quantum computers

81
00:03:16,680 --> 00:03:18,139
and artificial intelligence

82
00:03:18,760 --> 00:03:20,060
to boost the performance

83
00:03:20,520 --> 00:03:21,500
of these simulations.

84
00:03:23,000 --> 00:03:23,500
Recently,

85
00:03:23,800 --> 00:03:24,860
I was at Canada's

86
00:03:25,240 --> 00:03:26,379
Perimeter Institute

87
00:03:27,000 --> 00:03:28,620
and spoke to Javier

88
00:03:29,000 --> 00:03:29,500
Toledo

89
00:03:29,879 --> 00:03:30,379
Marine

90
00:03:30,885 --> 00:03:33,305
about the use of quantum assisted

91
00:03:33,844 --> 00:03:35,064
generative models

92
00:03:35,365 --> 00:03:36,745
in particle physics.

93
00:03:37,525 --> 00:03:39,784
Javier is based at the Triumph

94
00:03:40,165 --> 00:03:41,224
Particle Accelerator

95
00:03:41,605 --> 00:03:43,224
Center in British Columbia,

96
00:03:43,750 --> 00:03:46,969
and he also has an appointment at Perimeter.

97
00:03:47,590 --> 00:03:48,889
Here's that conversation.

98
00:03:57,235 --> 00:03:59,974
So I'm at the Perimeter Institute in Waterloo,

99
00:04:00,034 --> 00:04:03,014
Ontario, and I'm joined by Javier Toledo

100
00:04:03,314 --> 00:04:03,814
Marine.

101
00:04:04,275 --> 00:04:06,455
Hi, Javier. Welcome to the podcast.

102
00:04:06,995 --> 00:04:08,614
Hey, Hamish. Thank you.

103
00:04:09,330 --> 00:04:12,129
So Javier, you've recently published a paper that

104
00:04:12,129 --> 00:04:15,509
describes how quantum assisted artificial intelligence

105
00:04:16,129 --> 00:04:18,769
can be used to simulate how high energy

106
00:04:18,769 --> 00:04:19,269
particles

107
00:04:19,889 --> 00:04:21,189
interact with detectors

108
00:04:21,704 --> 00:04:23,644
such as ATLAS and CMS

109
00:04:24,105 --> 00:04:25,805
on the Large Hadron Collider.

110
00:04:26,584 --> 00:04:29,324
Why do physicists want to do these calculations,

111
00:04:29,785 --> 00:04:32,444
and why can't they be done using conventional

112
00:04:32,584 --> 00:04:33,084
computers?

113
00:04:33,544 --> 00:04:35,544
Yeah. So that's a great question. That's, I

114
00:04:35,544 --> 00:04:37,084
mean, that's the starting point.

115
00:04:37,649 --> 00:04:38,149
So

116
00:04:38,610 --> 00:04:39,830
as you might know,

117
00:04:40,449 --> 00:04:43,410
the Large Hadron Collider is closing. It's shutting

118
00:04:43,410 --> 00:04:43,910
down,

119
00:04:44,449 --> 00:04:47,430
for a few for some time now,

120
00:04:47,810 --> 00:04:50,629
because they're updating upgrading some of their detectors.

121
00:04:51,795 --> 00:04:53,795
So, so the new run of the LHC

122
00:04:53,795 --> 00:04:55,735
is called the high luminosity LHC.

123
00:04:56,115 --> 00:04:57,814
And so it's called high luminosity

124
00:04:58,115 --> 00:05:01,254
because, fundamentally, they're increasing the collision rate.

125
00:05:02,915 --> 00:05:06,250
So it's being increased by tenfold, I believe.

126
00:05:06,389 --> 00:05:08,550
And so that means that when they do

127
00:05:08,550 --> 00:05:11,209
these runs, we'll be getting more data. Right?

128
00:05:11,669 --> 00:05:13,990
And so this data is important because if

129
00:05:13,990 --> 00:05:15,129
you're trying to

130
00:05:15,589 --> 00:05:17,529
look at rare events,

131
00:05:18,375 --> 00:05:21,675
like, for instance, the dye Higgs, self interaction

132
00:05:21,814 --> 00:05:22,314
prompt,

133
00:05:23,334 --> 00:05:25,334
well, you actually have to it turns out

134
00:05:25,334 --> 00:05:27,354
that the dye Higgs is 1,000

135
00:05:27,735 --> 00:05:30,074
times less often than just the Higgs.

136
00:05:30,694 --> 00:05:33,149
So you need a lot more data, and

137
00:05:33,149 --> 00:05:35,470
so that's why it's being upgraded. And on

138
00:05:35,470 --> 00:05:36,930
the simulation side,

139
00:05:37,230 --> 00:05:38,290
you also need

140
00:05:38,750 --> 00:05:40,209
similar amount of data,

141
00:05:40,750 --> 00:05:42,769
and so that is becoming increasingly

142
00:05:43,149 --> 00:05:43,649
challenging.

143
00:05:44,110 --> 00:05:46,129
And one of the reasons why it's challenging

144
00:05:46,350 --> 00:05:47,730
is because of the

145
00:05:48,254 --> 00:05:48,754
calorimeter

146
00:05:49,134 --> 00:05:50,435
pipeline in the simulation.

147
00:05:51,134 --> 00:05:53,555
So there's in the ATLAS

148
00:05:54,014 --> 00:05:57,694
detect the ATLAS experiment, there's a detector, which

149
00:05:57,694 --> 00:05:58,514
is the calorimeter.

150
00:05:58,894 --> 00:06:01,235
And so when particles go through this calorimeter,

151
00:06:01,949 --> 00:06:03,490
they deposit the energy,

152
00:06:03,870 --> 00:06:06,189
but they also create secondary particles. And so

153
00:06:06,189 --> 00:06:07,810
they create these these showers.

154
00:06:08,509 --> 00:06:10,689
And so when that happens in the simulation,

155
00:06:10,829 --> 00:06:12,910
you need to track all of these particles

156
00:06:12,910 --> 00:06:14,449
as they go through the detector.

157
00:06:14,774 --> 00:06:16,375
And so you can imagine that if you

158
00:06:16,375 --> 00:06:18,535
increase the number of particles going through that,

159
00:06:18,535 --> 00:06:20,074
then you're going

160
00:06:20,615 --> 00:06:23,035
to need be needing more data from that.

161
00:06:23,654 --> 00:06:26,455
And so that is where it's becoming challenging

162
00:06:26,455 --> 00:06:27,595
because with the current

163
00:06:28,360 --> 00:06:30,939
computational capabilities, it turns out that it's

164
00:06:31,800 --> 00:06:32,300
unsurmountable

165
00:06:32,600 --> 00:06:35,660
at this point, using the traditional methods.

166
00:06:36,600 --> 00:06:38,759
So that is why people around the world,

167
00:06:38,759 --> 00:06:40,939
scientists, they're looking to use,

168
00:06:41,480 --> 00:06:44,194
different tools like deep generative models

169
00:06:44,574 --> 00:06:46,194
to generate these showers.

170
00:06:46,735 --> 00:06:49,235
So in our case, we're combining deep generative

171
00:06:49,295 --> 00:06:50,175
models with,

172
00:06:50,814 --> 00:06:53,634
quantum computing or, in this case, quantum annealers.

173
00:06:54,654 --> 00:06:57,160
So I think that answers the question. Right.

174
00:06:57,160 --> 00:06:59,160
So so the idea is that you're you're

175
00:06:59,160 --> 00:07:00,779
gonna have so many collisions

176
00:07:01,400 --> 00:07:03,740
going on in these detectors

177
00:07:04,360 --> 00:07:04,860
that,

178
00:07:05,319 --> 00:07:07,960
and you need to or I I suppose

179
00:07:07,960 --> 00:07:10,199
the physicists at CERN will need to focus

180
00:07:10,199 --> 00:07:11,764
in on a specific,

181
00:07:12,944 --> 00:07:15,745
event. But, of course, they they probably don't

182
00:07:15,745 --> 00:07:18,324
know exactly what that event will look like,

183
00:07:18,384 --> 00:07:20,785
and that's where your research comes in. You

184
00:07:20,865 --> 00:07:23,345
you're you do your simulations and you say,

185
00:07:23,345 --> 00:07:24,564
well, we think

186
00:07:24,889 --> 00:07:26,430
the event you're looking for

187
00:07:26,810 --> 00:07:28,970
is going to look like this in the

188
00:07:28,970 --> 00:07:31,709
detector. Yeah. That that's correct. So so basically,

189
00:07:31,769 --> 00:07:33,769
like, you have your experiment and the way

190
00:07:33,769 --> 00:07:35,850
to validate your experiment with theory, you do

191
00:07:35,850 --> 00:07:38,669
these simulations and then you compare your hypothesis

192
00:07:38,970 --> 00:07:41,435
from your simulations with what you actually get

193
00:07:41,435 --> 00:07:44,495
from from this experiment. Correct? I see. Okay.

194
00:07:45,914 --> 00:07:49,774
And can you explain in relatively simple terms,

195
00:07:50,074 --> 00:07:51,534
I mean, that might not be

196
00:07:51,834 --> 00:07:53,615
that might not be an easy task,

197
00:07:54,089 --> 00:07:54,589
How,

198
00:07:55,209 --> 00:07:56,589
your quantum assisted

199
00:07:56,889 --> 00:07:57,389
generative

200
00:07:57,930 --> 00:07:58,910
model works?

201
00:08:00,490 --> 00:08:03,209
Sure. Yeah. That's, okay. So let me give

202
00:08:03,209 --> 00:08:04,029
it a try.

203
00:08:04,970 --> 00:08:06,655
So, I mean, let let me start with

204
00:08:06,655 --> 00:08:08,995
generative models. So generative models, they're,

205
00:08:09,694 --> 00:08:11,694
way older than the current revolution that we're

206
00:08:11,694 --> 00:08:13,235
living with, with AI.

207
00:08:14,014 --> 00:08:16,415
And you can think about it as the

208
00:08:16,814 --> 00:08:18,814
like, they come to solve the task of

209
00:08:18,814 --> 00:08:19,314
generating

210
00:08:20,360 --> 00:08:24,220
data, generating samples from specific probability distributions.

211
00:08:24,600 --> 00:08:27,899
Right? So sometimes, like, many years ago when,

212
00:08:28,360 --> 00:08:31,500
you didn't have all these, programming languages, like,

213
00:08:31,639 --> 00:08:34,759
when you wanted to generate, Gaussian distributed random

214
00:08:34,759 --> 00:08:35,259
numbers,

215
00:08:35,815 --> 00:08:37,274
you would normally use,

216
00:08:37,654 --> 00:08:39,975
the Box Mueller method in which you would

217
00:08:39,975 --> 00:08:42,855
generate uniformly distributed random numbers, and then you

218
00:08:42,855 --> 00:08:45,834
would give those random numbers to a function,

219
00:08:46,054 --> 00:08:48,074
and then you would get your Gaussian distributed

220
00:08:48,294 --> 00:08:49,080
random numbers.

221
00:08:49,639 --> 00:08:51,240
These days, you you I mean, you just

222
00:08:51,240 --> 00:08:53,639
use your whatever. If you're using Python, it

223
00:08:53,639 --> 00:08:55,799
already has a a package that does that

224
00:08:55,799 --> 00:08:57,960
for you. So in that way, it's like

225
00:08:57,960 --> 00:09:00,040
we're trying to generate what we're working on

226
00:09:00,040 --> 00:09:03,160
is generating these showers, which come from a

227
00:09:03,160 --> 00:09:04,299
probability distribution,

228
00:09:04,695 --> 00:09:06,934
but we don't know that probability distribution a

229
00:09:06,934 --> 00:09:07,434
priori

230
00:09:08,134 --> 00:09:09,975
rate. So what we're doing is that we're

231
00:09:09,975 --> 00:09:12,075
sampling from an easier distribution,

232
00:09:13,014 --> 00:09:15,174
and then we pass it through some function.

233
00:09:15,174 --> 00:09:16,855
In this case, this function is a neural

234
00:09:16,855 --> 00:09:19,960
network, and then it gets converted to the

235
00:09:19,960 --> 00:09:20,460
shower.

236
00:09:20,840 --> 00:09:23,080
Now the way we are doing it is

237
00:09:23,080 --> 00:09:25,879
that we use quantum annealers as the initial

238
00:09:25,879 --> 00:09:26,379
sampler.

239
00:09:27,000 --> 00:09:29,160
So we sample from the quantum annealer. We

240
00:09:29,160 --> 00:09:31,340
get a a random vector.

241
00:09:31,664 --> 00:09:33,264
We pass it through a neural network, and

242
00:09:33,264 --> 00:09:34,164
we get the shower.

243
00:09:34,865 --> 00:09:36,865
Now that's the way once the model is

244
00:09:36,865 --> 00:09:39,105
trained, we we use it. Right? To train

245
00:09:39,105 --> 00:09:40,565
it, it's a bit more complicated.

246
00:09:41,345 --> 00:09:42,084
But, essentially,

247
00:09:42,464 --> 00:09:42,964
it's

248
00:09:43,424 --> 00:09:45,184
you can think about it as a data

249
00:09:45,184 --> 00:09:48,149
driven machine learning problem in which you have

250
00:09:48,210 --> 00:09:50,290
a data set that you feed to this

251
00:09:50,290 --> 00:09:52,629
model and you try to reconstruct it.

252
00:09:53,009 --> 00:09:55,570
And by reconstructing it, then you're you're, fine

253
00:09:55,570 --> 00:09:57,590
tuning these parameters in your neural network.

254
00:09:58,049 --> 00:10:00,129
And then you use this sampler, the quantum

255
00:10:00,129 --> 00:10:01,889
analyzer, to to sample from it and pass

256
00:10:01,889 --> 00:10:03,304
it through the through the neural

257
00:10:03,684 --> 00:10:05,284
network. I see. And and the data that

258
00:10:05,284 --> 00:10:06,504
you're using, is that

259
00:10:06,965 --> 00:10:07,784
data from

260
00:10:08,164 --> 00:10:10,264
scattering experiments that have been done

261
00:10:10,725 --> 00:10:11,465
in detectors?

262
00:10:12,004 --> 00:10:12,504
Or

263
00:10:13,125 --> 00:10:14,264
do you also incorporate

264
00:10:14,884 --> 00:10:15,384
physics

265
00:10:15,759 --> 00:10:17,519
into that as well, or is it purely

266
00:10:17,519 --> 00:10:19,919
just the the data, or is it both,

267
00:10:19,919 --> 00:10:20,579
I suppose?

268
00:10:21,360 --> 00:10:23,679
So we started off by using this dataset

269
00:10:23,679 --> 00:10:24,899
called the Cayo Challenge.

270
00:10:25,759 --> 00:10:27,839
And so the Cayo Challenge is a challenge

271
00:10:27,839 --> 00:10:30,914
that started, it was launched in 2022,

272
00:10:31,394 --> 00:10:33,375
by a group of,

273
00:10:33,914 --> 00:10:36,834
scientists around the world. And so the the

274
00:10:36,834 --> 00:10:38,674
scope of this challenge is precisely to see

275
00:10:38,674 --> 00:10:40,995
if we can use generative models to,

276
00:10:41,794 --> 00:10:44,399
to to sample, to generate these showers.

277
00:10:44,860 --> 00:10:47,120
So it's a publicly available dataset,

278
00:10:47,740 --> 00:10:49,200
and it's a

279
00:10:49,899 --> 00:10:52,540
different dataset than what you would normally get

280
00:10:52,540 --> 00:10:53,679
from the simulation

281
00:10:54,059 --> 00:10:56,475
pipeline that they used in in ATLAS. Right?

282
00:10:56,554 --> 00:10:59,754
It's still generated from that simulation pipeline, but

283
00:10:59,754 --> 00:11:03,434
it's then processed in a way that you

284
00:11:03,434 --> 00:11:05,754
might say it makes it maybe easier for

285
00:11:05,754 --> 00:11:07,674
the neural network to learn it. Right? That

286
00:11:07,834 --> 00:11:09,269
I mean, very yeah.

287
00:11:10,470 --> 00:11:12,470
So that's the dataset that we used in

288
00:11:12,470 --> 00:11:13,529
this this paper.

289
00:11:14,149 --> 00:11:18,149
And so right now, we're now looking to

290
00:11:18,149 --> 00:11:18,649
use

291
00:11:19,509 --> 00:11:22,330
data that's actually generated for the ATLAS experiment,

292
00:11:22,389 --> 00:11:25,654
more realistic dataset. Right? Now this dataset is

293
00:11:25,654 --> 00:11:27,975
composed by like, you think think about it

294
00:11:27,975 --> 00:11:30,294
this way. So when the collision happens, you

295
00:11:30,294 --> 00:11:30,794
get,

296
00:11:31,575 --> 00:11:33,415
particles going through that. Right? And then these

297
00:11:33,415 --> 00:11:35,415
particles go through the detectors. Right? So the

298
00:11:35,415 --> 00:11:37,575
data that we're feeding through to to our

299
00:11:37,575 --> 00:11:39,720
model, it's the data in which you you

300
00:11:39,720 --> 00:11:43,100
have that an incident particle entering the calorimeter

301
00:11:43,480 --> 00:11:45,019
with a specific energy

302
00:11:45,399 --> 00:11:47,480
and a specific type. Right? So a specific

303
00:11:47,480 --> 00:11:49,899
energy in the order of of GeV,

304
00:11:50,440 --> 00:11:52,945
giga electron volts. And the type of particle

305
00:11:52,945 --> 00:11:56,384
that we've considered thus far is, electrons. Right?

306
00:11:56,384 --> 00:11:58,165
But we are also looking to consider,

307
00:11:59,024 --> 00:12:02,004
photons and pions in the near future.

308
00:12:02,625 --> 00:12:03,285
I see.

309
00:12:03,679 --> 00:12:06,579
So Javier, you've used a a quantum annealer

310
00:12:07,039 --> 00:12:08,740
from D Wave systems.

311
00:12:09,279 --> 00:12:12,000
What is quantum annealing? And how does it

312
00:12:12,000 --> 00:12:15,279
differ from the the gate based models of

313
00:12:15,279 --> 00:12:17,024
quantum computation that

314
00:12:17,425 --> 00:12:19,264
maybe some of our listeners will be more

315
00:12:19,264 --> 00:12:20,165
familiar with.

316
00:12:21,184 --> 00:12:22,644
Right. Yeah. So then

317
00:12:23,345 --> 00:12:25,045
okay. So you have

318
00:12:25,665 --> 00:12:28,945
basically two different types of quantum computing, which

319
00:12:28,945 --> 00:12:30,725
is, like you mentioned, gate based

320
00:12:31,160 --> 00:12:33,019
and adiabatic quantum computing.

321
00:12:34,120 --> 00:12:36,940
So in theory, both of them are universal,

322
00:12:37,480 --> 00:12:39,420
quantum tuning in the sense that you can

323
00:12:39,960 --> 00:12:43,240
do any quantum algorithm and and using both

324
00:12:43,240 --> 00:12:43,740
technologies.

325
00:12:44,575 --> 00:12:47,294
In practice, there are nuances. You have, like,

326
00:12:47,294 --> 00:12:50,174
decoherence. You have noise in it in both

327
00:12:50,174 --> 00:12:50,674
cases.

328
00:12:51,214 --> 00:12:53,615
And so in the case of gate based

329
00:12:53,615 --> 00:12:54,595
quantum computing,

330
00:12:55,375 --> 00:12:58,495
what you're using there are unitary gates that

331
00:12:58,495 --> 00:13:00,115
operate on the qubits.

332
00:13:00,550 --> 00:13:02,950
So you can think about it more like

333
00:13:02,950 --> 00:13:04,490
a in a digital way.

334
00:13:04,950 --> 00:13:07,290
Whereas adiabatic quantum computing, you're

335
00:13:08,070 --> 00:13:08,570
continuously

336
00:13:09,110 --> 00:13:12,790
evolving your quantum system with, what's called an

337
00:13:12,790 --> 00:13:14,090
evolution operator.

338
00:13:14,434 --> 00:13:16,054
You're you're evolving your system.

339
00:13:16,914 --> 00:13:20,514
And what this adiabatic quantum computing does, it

340
00:13:20,514 --> 00:13:24,214
finds the ground state of a specific Hamiltonian

341
00:13:24,434 --> 00:13:25,735
that encodes your problem.

342
00:13:26,514 --> 00:13:28,500
Now like I said, in practice, there are

343
00:13:28,740 --> 00:13:29,639
lots of nuances

344
00:13:30,340 --> 00:13:30,840
related

345
00:13:31,299 --> 00:13:32,039
to decoherence

346
00:13:32,500 --> 00:13:35,220
and noise and error correction and and error

347
00:13:35,220 --> 00:13:36,519
in your in your system.

348
00:13:37,539 --> 00:13:39,620
So in our case, we're not when we

349
00:13:39,620 --> 00:13:42,019
evolve this simultaneous to get the the ground

350
00:13:42,019 --> 00:13:43,779
state, we're not really interested in the ground

351
00:13:43,779 --> 00:13:44,235
state.

352
00:13:44,875 --> 00:13:46,654
We're interested in getting a

353
00:13:47,115 --> 00:13:49,375
sample that comes from a Boltzmann distribution.

354
00:13:49,995 --> 00:13:51,695
So it just happens that

355
00:13:52,714 --> 00:13:54,014
this quantum annealer,

356
00:13:54,394 --> 00:13:55,115
when you,

357
00:13:55,514 --> 00:13:57,534
when you do the annealing process

358
00:13:58,079 --> 00:14:01,360
in a specific time window, you're going to

359
00:14:01,360 --> 00:14:03,360
get a sample that comes from a Boltzmann

360
00:14:03,360 --> 00:14:06,399
distribution. Right? And so we use that because,

361
00:14:06,399 --> 00:14:08,879
basically, our framework is a combination of what's

362
00:14:08,879 --> 00:14:10,500
called a variational autoencoder

363
00:14:11,164 --> 00:14:13,584
and what's called a restricted Boltzmann machine.

364
00:14:14,044 --> 00:14:16,784
So this restricted Boltzmann machine is,

365
00:14:17,404 --> 00:14:19,644
like I mentioned earlier, where we sample from.

366
00:14:19,644 --> 00:14:21,024
We use it as a sampler.

367
00:14:21,564 --> 00:14:23,345
And and so that makes it,

368
00:14:24,049 --> 00:14:27,169
oh, like, basically a direct way to sample

369
00:14:27,169 --> 00:14:29,669
from a quantum annealer. Like, it's completely straightforward,

370
00:14:30,129 --> 00:14:31,750
sampling from a quantum annealer,

371
00:14:32,209 --> 00:14:34,370
and using it as a as a quantum

372
00:14:34,370 --> 00:14:36,850
version of a restricted Bose machine. I see.

373
00:14:36,850 --> 00:14:38,850
And so so is the is the idea

374
00:14:38,850 --> 00:14:41,355
that, you know, you've you've this electron has

375
00:14:41,355 --> 00:14:42,654
gone into the detector,

376
00:14:43,034 --> 00:14:44,414
and it's going to interact.

377
00:14:44,794 --> 00:14:46,414
And there's lots of different

378
00:14:47,195 --> 00:14:48,495
things that can happen.

379
00:14:48,955 --> 00:14:51,355
You know, this interaction can occur, then that

380
00:14:51,355 --> 00:14:54,090
interaction can occur. And, you know, it probably

381
00:14:54,090 --> 00:14:55,309
just blows up into

382
00:14:55,850 --> 00:14:56,430
a huge,

383
00:14:57,370 --> 00:14:59,629
I suppose, space of possible things.

384
00:15:00,170 --> 00:15:02,170
Do is the idea of the of the

385
00:15:02,170 --> 00:15:02,670
quantum

386
00:15:03,050 --> 00:15:06,269
annealer, does it give you the most probable,

387
00:15:08,100 --> 00:15:08,595
outcome

388
00:15:08,995 --> 00:15:10,534
of of that collision

389
00:15:11,394 --> 00:15:13,554
process? Is is that what you're looking for,

390
00:15:13,554 --> 00:15:14,914
or is it is it a bit more

391
00:15:14,914 --> 00:15:16,134
complicated than that?

392
00:15:16,754 --> 00:15:18,615
Well, I would say that it's not necessarily

393
00:15:18,754 --> 00:15:19,735
the most probable,

394
00:15:20,754 --> 00:15:22,929
event that we're looking for. We're just looking

395
00:15:22,929 --> 00:15:24,789
to be able to sample from

396
00:15:25,570 --> 00:15:27,730
from the manifold where these like, when when

397
00:15:27,730 --> 00:15:29,649
you mentioned this, you can think about, like,

398
00:15:29,649 --> 00:15:31,889
when this electron goes through this detector and

399
00:15:31,889 --> 00:15:33,029
generates the shower

400
00:15:33,490 --> 00:15:34,789
so that event,

401
00:15:36,004 --> 00:15:37,625
falls in some kind of manifold.

402
00:15:38,325 --> 00:15:40,004
And so we want to be able to

403
00:15:40,004 --> 00:15:42,485
sample from from those manifold from from that

404
00:15:42,485 --> 00:15:43,464
manifold. Right?

405
00:15:43,764 --> 00:15:46,884
And so this, this quantum annealer, this this

406
00:15:46,884 --> 00:15:49,620
framework of combining the quantum annealer with the

407
00:15:49,860 --> 00:15:50,839
variational encoder

408
00:15:51,779 --> 00:15:54,360
allows us to efficiently sample

409
00:15:54,740 --> 00:15:57,860
from that, from that manifold. Right? And so

410
00:15:57,860 --> 00:16:00,899
when I say efficiently, I'm thinking of mainly

411
00:16:00,899 --> 00:16:02,625
two things. One is

412
00:16:03,184 --> 00:16:04,325
accurately capturing

413
00:16:05,184 --> 00:16:07,504
the the physics or the properties of these

414
00:16:07,504 --> 00:16:08,325
these showers,

415
00:16:09,424 --> 00:16:12,384
for which we we we have metrics for

416
00:16:12,384 --> 00:16:14,485
that, and we can compare our

417
00:16:14,949 --> 00:16:17,350
framework with other frameworks, and we're doing well

418
00:16:17,350 --> 00:16:19,190
on that front. But, also, you wanna do

419
00:16:19,190 --> 00:16:21,350
it fast. Right? Because in the end, that's

420
00:16:21,350 --> 00:16:23,029
the the problem that we're tackling that. Can

421
00:16:23,029 --> 00:16:24,250
we generate those showers,

422
00:16:24,709 --> 00:16:27,225
and can we do it faster than, like,

423
00:16:27,384 --> 00:16:27,884
current,

424
00:16:28,664 --> 00:16:31,324
traditional methods are able to do so? Right.

425
00:16:31,544 --> 00:16:34,024
And the the I mean, these the the

426
00:16:34,024 --> 00:16:36,345
way that a particle will interact with a

427
00:16:36,345 --> 00:16:38,204
detector, that's, I mean, that's

428
00:16:38,904 --> 00:16:40,605
governed by quantum mechanics.

429
00:16:41,480 --> 00:16:43,740
You know, it's all probabilities, etcetera.

430
00:16:44,279 --> 00:16:46,679
Does the fact that you're using a quantum

431
00:16:46,679 --> 00:16:47,179
annealer,

432
00:16:47,720 --> 00:16:50,360
does that help you go faster in the

433
00:16:50,360 --> 00:16:52,120
sense that, you know, the system you're trying

434
00:16:52,120 --> 00:16:53,419
to simulate is quantum,

435
00:16:53,879 --> 00:16:55,980
your hardware is quantum?

436
00:16:56,725 --> 00:16:58,725
Therefore, you know, that that I suppose the

437
00:16:58,725 --> 00:17:00,264
old argument of Richard Feynman

438
00:17:00,644 --> 00:17:03,205
about, you know, why quantum computers could be

439
00:17:03,205 --> 00:17:06,724
useful for studying quantum systems. Does that apply

440
00:17:06,724 --> 00:17:07,224
here?

441
00:17:08,819 --> 00:17:10,919
Yeah. That's a very interesting question.

442
00:17:12,099 --> 00:17:14,339
I would say that at this point, it

443
00:17:14,339 --> 00:17:16,579
doesn't apply because we still like, when we

444
00:17:16,579 --> 00:17:18,119
sample from the quantum annealer,

445
00:17:19,220 --> 00:17:19,720
we're,

446
00:17:20,019 --> 00:17:22,419
I mean, we're doing a classical measurement, and

447
00:17:22,419 --> 00:17:25,945
then we're passing that through a classical algorithm.

448
00:17:25,945 --> 00:17:27,164
Right? So

449
00:17:27,625 --> 00:17:29,465
I would say that it doesn't apply here

450
00:17:29,465 --> 00:17:31,484
at this point, but that's certainly,

451
00:17:32,105 --> 00:17:34,664
like, the the scope of this project. Right?

452
00:17:34,664 --> 00:17:36,744
Because at this point, the the training that

453
00:17:36,744 --> 00:17:37,289
we do,

454
00:17:37,850 --> 00:17:38,750
it's basically

455
00:17:39,049 --> 00:17:40,190
a classical training.

456
00:17:40,970 --> 00:17:44,509
But we're certainly, like, we're looking forward to,

457
00:17:45,450 --> 00:17:47,869
getting more from this these quantum annealers

458
00:17:48,250 --> 00:17:50,250
and and using them for for this type

459
00:17:50,250 --> 00:17:53,204
of physics problem. Yeah. I see. And what

460
00:17:53,204 --> 00:17:54,025
does success

461
00:17:54,404 --> 00:17:56,565
mean? How how have you shown that your

462
00:17:56,565 --> 00:17:58,345
technique is useful for predicting

463
00:17:58,804 --> 00:18:01,065
how particles interact with the detectors?

464
00:18:02,724 --> 00:18:04,345
So at this point, what we're

465
00:18:04,884 --> 00:18:06,884
what we're using are it's a set of

466
00:18:06,884 --> 00:18:09,839
metrics that come from the kalo challenge and

467
00:18:09,839 --> 00:18:13,059
also metrics that are typically used by,

468
00:18:13,839 --> 00:18:15,059
high energy physicists.

469
00:18:15,919 --> 00:18:18,879
And so what we're basically doing is that

470
00:18:18,879 --> 00:18:20,419
we're comparing the

471
00:18:20,744 --> 00:18:21,484
the performance

472
00:18:21,785 --> 00:18:24,424
of our framework with traditional with the traditional

473
00:18:24,424 --> 00:18:27,484
method and also with other generative models. Right?

474
00:18:27,704 --> 00:18:28,444
So that's,

475
00:18:29,785 --> 00:18:32,505
how success looks like in at at this

476
00:18:32,505 --> 00:18:33,244
point. Right?

477
00:18:33,809 --> 00:18:34,549
By looking,

478
00:18:35,009 --> 00:18:37,089
how do we perform compared to to other

479
00:18:37,089 --> 00:18:39,169
methods? Right? And so at this at this

480
00:18:39,169 --> 00:18:39,669
stage,

481
00:18:40,049 --> 00:18:40,869
we're doing

482
00:18:41,409 --> 00:18:43,809
quite well in terms of quality and in

483
00:18:43,809 --> 00:18:44,549
terms of

484
00:18:45,089 --> 00:18:46,149
speed up. It's,

485
00:18:47,195 --> 00:18:48,954
as far as we we know, it's the

486
00:18:48,954 --> 00:18:50,095
the fastest method,

487
00:18:50,795 --> 00:18:53,035
compared to other What what what I mean,

488
00:18:53,035 --> 00:18:55,115
you mentioned speed up. What I mean, how

489
00:18:55,115 --> 00:18:57,055
much faster is your method than

490
00:18:57,434 --> 00:18:59,055
a method using a conventional

491
00:18:59,640 --> 00:19:00,140
supercomputer

492
00:19:00,440 --> 00:19:02,839
or what what whatever these calculations are done

493
00:19:02,839 --> 00:19:03,740
on at the moment?

494
00:19:04,839 --> 00:19:06,759
Right. Okay. So let me be nuanced here

495
00:19:06,759 --> 00:19:09,400
in this this question. So we've used the

496
00:19:09,559 --> 00:19:11,240
for the for the Caleb challenge, which is

497
00:19:11,240 --> 00:19:12,914
the dataset that we used for this,

498
00:19:13,555 --> 00:19:14,055
paper.

499
00:19:15,075 --> 00:19:16,295
And our results

500
00:19:16,755 --> 00:19:18,295
show that we're about

501
00:19:18,914 --> 00:19:22,434
1,000 times faster than the traditional method, which

502
00:19:22,434 --> 00:19:23,815
is called g e on four.

503
00:19:25,075 --> 00:19:26,055
Compared to

504
00:19:26,389 --> 00:19:28,950
other generative models that were part of the

505
00:19:28,950 --> 00:19:29,849
Cayla Challenge,

506
00:19:30,230 --> 00:19:31,049
we're also

507
00:19:32,950 --> 00:19:35,450
faster than than Dose Method, currently.

508
00:19:36,389 --> 00:19:38,710
I see. And are you are you at

509
00:19:38,710 --> 00:19:40,634
the point where you're sort of moving out

510
00:19:40,954 --> 00:19:44,315
from the research phase, let's say, to, you

511
00:19:44,315 --> 00:19:46,634
know, sort of letting this thing roll and

512
00:19:46,634 --> 00:19:50,315
doing calculations that you will then pass on

513
00:19:50,315 --> 00:19:52,650
to physicists working at CERN?

514
00:19:53,369 --> 00:19:55,369
Or is there more work to do before

515
00:19:55,369 --> 00:19:56,349
you can do that?

516
00:19:57,130 --> 00:19:57,630
Yeah.

517
00:19:58,170 --> 00:20:00,330
Great question. So that there's still work to

518
00:20:00,330 --> 00:20:01,869
be done because, like I mentioned,

519
00:20:02,250 --> 00:20:02,730
we've,

520
00:20:03,690 --> 00:20:05,690
we first used the Kalot challenge dataset, and

521
00:20:05,690 --> 00:20:08,349
now we're moving to ATLAS generated dataset.

522
00:20:08,865 --> 00:20:10,785
So now, like, we what we have to

523
00:20:10,785 --> 00:20:11,444
do is

524
00:20:11,904 --> 00:20:14,224
everything we showed for the kilo challenge, we

525
00:20:14,224 --> 00:20:15,444
now need to be

526
00:20:16,065 --> 00:20:17,984
we now need to show it for this

527
00:20:17,984 --> 00:20:19,125
ATLAS dataset.

528
00:20:20,144 --> 00:20:22,080
So that means we need

529
00:20:22,460 --> 00:20:23,600
to train models

530
00:20:24,220 --> 00:20:25,440
on multiple datasets,

531
00:20:26,140 --> 00:20:29,279
run these benchmarks, and show that we're actually

532
00:20:29,660 --> 00:20:31,740
as good as we should for the kill

533
00:20:31,740 --> 00:20:34,320
challenge. So I think that's still gonna take,

534
00:20:35,534 --> 00:20:38,194
a reason about a reasonable amount of time,

535
00:20:38,654 --> 00:20:39,714
but we're certainly

536
00:20:40,494 --> 00:20:41,794
looking forward to

537
00:20:42,335 --> 00:20:44,414
to pushing this to to a place in

538
00:20:44,414 --> 00:20:46,815
which it could be deployed. Right? Deployable, at

539
00:20:46,815 --> 00:20:47,315
least.

540
00:20:47,615 --> 00:20:49,474
And final question, Javier.

541
00:20:50,180 --> 00:20:52,119
Are there any sort of non

542
00:20:52,980 --> 00:20:53,960
particle physics

543
00:20:54,259 --> 00:20:55,960
applications to this technique?

544
00:20:56,500 --> 00:20:57,779
I mean, it could could it be used

545
00:20:57,779 --> 00:21:00,500
in other parts of physics or you useful

546
00:21:00,500 --> 00:21:03,220
to physicists working in condensed matter or, I

547
00:21:03,220 --> 00:21:04,440
don't know, even biologists,

548
00:21:05,054 --> 00:21:05,554
chemists,

549
00:21:05,855 --> 00:21:08,254
or I mean even, I don't know, a

550
00:21:08,254 --> 00:21:11,294
trucking company that wants to, you know, plan

551
00:21:11,294 --> 00:21:13,154
the most efficient route across,

552
00:21:13,774 --> 00:21:16,014
across Canada. Are there are there up other

553
00:21:16,014 --> 00:21:17,875
applications for this, or is it only

554
00:21:18,190 --> 00:21:20,049
sort of a particle physics thing?

555
00:21:20,750 --> 00:21:22,990
Right. Okay. So the the what we call

556
00:21:22,990 --> 00:21:25,549
the Cayo QVAE framework, which is the VAE

557
00:21:25,549 --> 00:21:27,950
with the using the the quantum annealer as

558
00:21:27,950 --> 00:21:28,609
a sampler,

559
00:21:28,910 --> 00:21:31,789
that you that you can use it for

560
00:21:31,789 --> 00:21:32,289
any

561
00:21:32,865 --> 00:21:35,365
machine learning problem generative model. Right?

562
00:21:35,664 --> 00:21:37,845
So you can use it either to generate

563
00:21:38,545 --> 00:21:39,045
pictures

564
00:21:39,744 --> 00:21:43,045
of people or of dogs or whatnot. Right?

565
00:21:43,664 --> 00:21:46,410
So it's not limited to to just physics.

566
00:21:46,410 --> 00:21:46,910
Right?

567
00:21:48,090 --> 00:21:50,090
However, you did mention, right, can you use

568
00:21:50,090 --> 00:21:52,830
it to optimize the routes of,

569
00:21:53,610 --> 00:21:56,250
no, like, delivery stuff like that. Right? So

570
00:21:56,250 --> 00:21:58,330
I guess that's a a good question. Right?

571
00:21:58,330 --> 00:22:00,430
So so quantum annealers, the

572
00:22:01,565 --> 00:22:03,884
the motivation is to to find this ground

573
00:22:03,884 --> 00:22:06,144
state. Right? So there are some problems

574
00:22:06,765 --> 00:22:08,924
that in principle you can map, like problems

575
00:22:08,924 --> 00:22:11,404
of optimization of routes that you can map

576
00:22:11,404 --> 00:22:13,750
to a quantum annealer. Right? So in that

577
00:22:13,750 --> 00:22:14,250
case,

578
00:22:14,950 --> 00:22:17,910
you wouldn't actually need the variational encoder. You

579
00:22:17,910 --> 00:22:20,730
would just use the the quantum annealer to

580
00:22:21,029 --> 00:22:24,150
to find that that, ground state of that

581
00:22:24,150 --> 00:22:24,650
optimum

582
00:22:25,029 --> 00:22:25,529
solution.

583
00:22:26,545 --> 00:22:28,785
That's the, I suppose, is that the famous

584
00:22:28,785 --> 00:22:29,765
traveling salesman

585
00:22:30,144 --> 00:22:32,384
problem, which I think is really difficult, isn't

586
00:22:32,384 --> 00:22:33,525
it? That's correct.

587
00:22:33,984 --> 00:22:34,484
Yeah.

588
00:22:35,424 --> 00:22:38,325
Yeah. That's correct. That's, that's the the salesperson,

589
00:22:39,105 --> 00:22:40,005
traveling problem.

590
00:22:41,710 --> 00:22:42,369
Of course,

591
00:22:42,829 --> 00:22:45,069
there you might have some extra constraints, like,

592
00:22:45,069 --> 00:22:46,829
in a real problem. Right? And that's where

593
00:22:46,829 --> 00:22:47,490
it gets

594
00:22:47,950 --> 00:22:50,829
really complicated. How how do you embed those

595
00:22:50,829 --> 00:22:51,329
constraints

596
00:22:51,869 --> 00:22:54,369
into the quantum in either? Mhmm. Right?

597
00:22:55,464 --> 00:22:57,704
Well, thanks so much, Javier. Thanks for speaking

598
00:22:57,704 --> 00:22:59,944
to Physics World about your research, and we

599
00:22:59,944 --> 00:23:01,944
look forward to, to hearing more from you

600
00:23:01,944 --> 00:23:02,764
and your colleagues.

601
00:23:03,384 --> 00:23:04,444
Thank you so much.

602
00:23:12,230 --> 00:23:14,730
That was Javier Toledo Marine

603
00:23:15,109 --> 00:23:19,130
in conversation with me at Canada's Perimeter Institute.

604
00:23:19,990 --> 00:23:22,970
The paper we spoke about is called Conditioned

605
00:23:23,744 --> 00:23:24,884
quantum assisted

606
00:23:25,424 --> 00:23:26,724
deep generative

607
00:23:27,424 --> 00:23:27,924
surrogate

608
00:23:28,305 --> 00:23:29,204
for particle

609
00:23:29,744 --> 00:23:30,244
calorimeter

610
00:23:31,025 --> 00:23:31,525
interactions,

611
00:23:32,224 --> 00:23:34,404
and it's published in NPJ

612
00:23:35,105 --> 00:23:36,325
Quantum Information.

613
00:23:37,099 --> 00:23:39,180
I'll put a link to the paper in

614
00:23:39,180 --> 00:23:40,480
the podcast notes.

615
00:23:41,340 --> 00:23:43,980
I'd like to thank Delft Circuits for their

616
00:23:43,980 --> 00:23:46,000
generous support of this episode,

617
00:23:46,619 --> 00:23:48,720
and also thanks to Javier

618
00:23:49,099 --> 00:23:49,599
Toledo

619
00:23:49,900 --> 00:23:50,400
Marine

620
00:23:50,714 --> 00:23:52,095
for joining me today.

621
00:23:52,474 --> 00:23:54,494
And as always, our producer,

622
00:23:54,875 --> 00:23:56,015
Fred Ailes.

623
00:23:56,714 --> 00:23:59,034
I'll be back again next week when I'm

624
00:23:59,034 --> 00:24:01,534
joined by a physicist and a sculptor

625
00:24:01,994 --> 00:24:04,875
to talk about the science and art of

626
00:24:04,875 --> 00:24:06,174
quantum steampunk.

627
00:24:07,039 --> 00:24:08,099
See you then.

628
00:24:14,319 --> 00:24:18,019
As gate based quantum computing continues to scale,

629
00:24:18,480 --> 00:24:21,380
Delft circuits provides the IO solutions

630
00:24:21,914 --> 00:24:23,215
that make it possible.