Slide for Arithmer Seminar given by Dr. Enrico Rinaldi (RIKEN) at Arithmer inc. The topic is how data science is used at frontier in research of physics, especially in high-energy and condensed matter physics. Actually, three month after from this seminar, Dr. Rinaldi joined Arithmer as AI researcher/engineer! "Arithmer Seminar" is weekly held, where professionals from within and outside our company give lectures on their respective expertise. The slides are made by the lecturer from outside our company, and shared here with his/her permission. Arithmer株式会社は東京大学大学院数理科学研究科発の数学の会社です。私達は現代数学を応用して、様々な分野のソリューションに、新しい高度AIシステムを導入しています。AIをいかに上手に使って仕事を効率化するか、そして人々の役に立つ結果を生み出すのか、それを考えるのが私たちの仕事です。 Arithmer began at the University of Tokyo Graduate School of Mathematical Sciences. Today, our research of modern mathematics and AI systems has the capability of providing solutions when dealing with tough complex issues. At Arithmer we believe it is our job to realize the functions of AI through improving work efficiency and producing more useful results for society.

1. 2 0 1 9 / 0 6 / 1 3
THE DATA SCIENCE OF PHYSICS
RIKEN ENRICO RINALDI
Arithmer Seminar」とは、週1回、弊社社員や外部から様々な分野の方を講師に招いて、社員の知見を広めることを目的としたセミナーです。
次頁以降のスライドは、講師を務めた外部の方の作成によるものであり、許可を得て公開しています。

2. THE DATA SCIENCE OF PHYSICS
ENRICO RINALDI
June 13, 2019 Arithmer Inc.RIKEN Nishina Center for Accelerator-based Science

3. THE DATA SCIENCE OF PHYSICS
ENRICO RINALDI
June 13, 2019 Arithmer Inc.
ACCELERATE PHYSICS DISCOVERIES WITH AI
RIKEN Nishina Center for Accelerator-based Science

4. PERSONAL WEBSITE: HTTPS://ERINALDI-1.NETLIFY.COM
WHO AM I?
▸ I am a theoretical particle physicist (High Energy and Nuclear)
▸ I am a computational physicist (High Performance Computing)
▸ I transform abstract concepts into practical problems:
▸ from a mathematical theory
▸ to an algorithm deployed on supercomputers
▸ analyzing large amounts of data to test hypothesis
▸ Driven by curiosity…following the scientific method

5. [Sandbox Studio, Chicago with Ana Kova]

6. [Sandbox Studio, Chicago with Ana Kova]
WHAT ARE WE MADE OF?
WHAT IS THE ORIGIN OF MATTER?
WHAT ARE THE LAWS OF THE
UNIVERSE?

7. [Sandbox Studio, Chicago with Ana Kova]
WHAT ARE WE MADE OF?
WHAT IS THE ORIGIN OF MATTER?
WHAT ARE THE LAWS OF THE
UNIVERSE?
higgstan.com

8. [Sandbox Studio, Chicago with Ana Kova]
WHAT ARE WE MADE OF?
WHAT IS THE ORIGIN OF MATTER?
WHAT ARE THE LAWS OF THE
UNIVERSE?
higgstan.com

9. ATLAS detector

10. ATLAS detector

11. ATLAS detector

12. Particle Physics “events”

13. Particle Physics “events”
Self-driving “events”

14. Particle “tracks” and “showers”

15. Particle “tracks” and “showers”
FUTURE UPGRADED EXPERIMENTS WILL PRODUCE ~10X MORE DATA!
HOW CAN WE MAKE SENSE OF IT AND HOW FAST?

16. [KIYOSHI TAKAHASE SEGUNDO/ALAMY STOCK PHOTO]
“AI PROMISES TO SUPERCHARGE
THE PROCESS OF DISCOVERY”
VOL 357 ISSUE 6346
The scientists’ apprentice - Tim Appenzeller

17. [KIYOSHI TAKAHASE SEGUNDO/ALAMY STOCK PHOTO]
“AI PROMISES TO SUPERCHARGE
THE PROCESS OF DISCOVERY”
VOL 357 ISSUE 6346
The scientists’ apprentice - Tim Appenzeller
web link

18. [KIYOSHI TAKAHASE SEGUNDO/ALAMY STOCK PHOTO]
“AI PROMISES TO SUPERCHARGE
THE PROCESS OF DISCOVERY”
VOL 357 ISSUE 6346
The scientists’ apprentice - Tim Appenzeller
web link

19. DIFFERENT LEARNING SCHEMES FOR A MACHINE
SUPERVISED UNSUPERVISED
*REINFORCEMENT

20. DIFFERENT LEARNING SCHEMES FOR A MACHINE
SUPERVISED UNSUPERVISED
*REINFORCEMENT
Data Labels

21. DIFFERENT LEARNING SCHEMES FOR A MACHINE
SUPERVISED UNSUPERVISED
*REINFORCEMENT
Data Labels
Albert Einstein
Stephen Hawking

22. DIFFERENT LEARNING SCHEMES FOR A MACHINE
SUPERVISED UNSUPERVISED
*REINFORCEMENT
Data Labels
Albert Einstein
Stephen Hawking
https://einstein.onrender.com

23. DIFFERENT LEARNING SCHEMES FOR A MACHINE
SUPERVISED UNSUPERVISED
*REINFORCEMENT
Data Labels
Albert Einstein
Stephen Hawking
https://einstein.onrender.com
Data ??

24. DIFFERENT LEARNING SCHEMES FOR A MACHINE
SUPERVISED UNSUPERVISED
*REINFORCEMENT
Data Labels
Albert Einstein
Stephen Hawking
https://einstein.onrender.com
Data ??
DeNA Co., Ltd., Tokyo, Japan

25. DIFFERENT LEARNING SCHEMES FOR A MACHINE
SUPERVISED UNSUPERVISED
*REINFORCEMENT
Data Labels
Albert Einstein
Stephen Hawking
https://einstein.onrender.com
ability to recognize, classify and characterize complex sets of data
Data ??
DeNA Co., Ltd., Tokyo, Japan

26. [Sandbox Studio, Chicago with Ana Kova]

27. [Sandbox Studio, Chicago with Ana Kova]

28. [Sandbox Studio, Chicago with Ana Kova]
SEMANTIC SEGMENTATION

29. [Sandbox Studio, Chicago with Ana Kova]
SEMANTIC SEGMENTATION

30. [Sandbox Studio, Chicago with Ana Kova]
SEMANTIC SEGMENTATION

31. [Sandbox Studio, Chicago with Ana Kova]
MASK
IMAGE

32. Scientific paper link
NETWORK ARCHITECTURE
deeplearnphysics organization

33. Scientific paper link
NETWORK ARCHITECTURE
INPUT “LABEL” OUTPUT
deeplearnphysics organization

34. Scientific paper link
NETWORK ARCHITECTURE
INPUT “LABEL” OUTPUT
deeplearnphysics organization

35. Scientific paper link
INPUT “LABEL” OUTPUT
REAL DATA!
deeplearnphysics organization

36. Scientific paper link
INPUT “LABEL” OUTPUT
REAL DATA!
deeplearnphysics organization
LABELED BY A
HUMAN!

37. Scientific paper link
INPUT “LABEL” OUTPUT
REAL DATA!
deeplearnphysics organization
LABELED BY A
HUMAN!
WE WANT THE
MACHINE TO
MAKE THIS
TASK FASTER

38. CONDENSED MATTER PHYSICS
THE ISING MODEL
H ({si}) = − J
∑
⟨i,j⟩
sisj−h
∑
i
si
1D
2D
s = + 1
s = − 1
SPIN
p({si}) =
e−βH({si})
Z
β =
1
kBT
Z =
∑
{si}
e−βH({si})
Probability of state:
Partition function:
Hamiltonian of spins with local interactions:
#{si} = 2N
N
Inverse temperature
Number of states:
E = ⟨H⟩ = Z−1
∑
{si}
H ({si}) e−βH({si})

39. CONDENSED MATTER PHYSICS
THE ISING MODEL
Model for:
‣ Magnetic materials
‣ Neuronal activity
‣ Quantum computers
H ({si}) = − J
∑
⟨i,j⟩
sisj−h
∑
i
si
1D
2D
s = + 1
s = − 1
SPIN
p({si}) =
e−βH({si})
Z
β =
1
kBT
Z =
∑
{si}
e−βH({si})
Probability of state:
Partition function:
Hamiltonian of spins with local interactions:
#{si} = 2N
N
Inverse temperature
Number of states:
E = ⟨H⟩ = Z−1
∑
{si}
H ({si}) e−βH({si})

40. CONDENSED MATTER PHYSICS
THE ISING MODEL
Model for:
‣ Magnetic materials
‣ Neuronal activity
‣ Quantum computers
H ({si}) = − J
∑
⟨i,j⟩
sisj−h
∑
i
si
1D
2D
s = + 1
s = − 1
SPIN
p({si}) =
e−βH({si})
Z
β =
1
kBT
Z =
∑
{si}
e−βH({si})
Probability of state:
Partition function:
Hamiltonian of spins with local interactions:
#{si} = 2N
N
Inverse temperature
Number of states:
E = ⟨H⟩ = Z−1
∑
{si}
H ({si}) e−βH({si})

41. PHASE TRANSITIONS
ORDERED AND DISORDERED PHASES
COLD HOT
Ising
Water
T

42. PHASE TRANSITIONS
ORDERED AND DISORDERED PHASES
COLD HOT
Ising
Water
T

43. PHASE TRANSITIONS
ORDERED AND DISORDERED PHASES
COLD HOT
Ising
Water
T
ORDERED

44. PHASE TRANSITIONS
ORDERED AND DISORDERED PHASES
COLD HOT
Ising
Water
T
ORDERED

45. PHASE TRANSITIONS
ORDERED AND DISORDERED PHASES
COLD HOT
Ising
Water
T
DISORDEREDORDERED

46. PHASE TRANSITIONS
ORDERED AND DISORDERED PHASES
COLD HOTCRITICAL
Ising
Water
T
PHASE TRANSITION DISORDEREDORDERED

47. PHASE TRANSITIONS
ORDERED AND DISORDERED PHASES
COLD HOTCRITICAL
Ising
Water
T
PHASE TRANSITION DISORDEREDORDERED
CAN WE PREDICT ANY OF THIS WITHOUT KNOWING THE
MICROSCOPIC DETAILS?
DO WE NEED A FULL UNDERSTANDING OF THE COMPLEX DYNAMICS
TO MAKE PREDICTIONS?

48. MACHINE LEARNING PHASES OF MATTER
COLD
HOT
Scientific paper link
Playground

49. MACHINE LEARNING PHASES OF MATTER
COLD
HOT
IMAGE/Data Labels
Scientific paper link
Playground

50. MACHINE LEARNING PHASES OF MATTER
COLD
HOT
??
IMAGE/Data Labels
SUPERVISED CLASSIFICATION
SVM, MLP, CNN
Scientific paper link
Playground

51. MACHINE LEARNING PHASES OF MATTER
COLD
HOT
IMAGE/Data Labels
Scientific paper link
Input layer
Hidden layer
Output layer
A
1.0 1.5 2.0 2.5 3.0 3.5
0.0
0.2
0.4
0.6
0.8
1.0
Outputlayer
B
L = 30 T < Tc
T > Tc
1.0 1.5 2.0 2.5 3.0 3.5
T
0.0
0.2
0.4
0.6
0.8
1.0
Accuracy
C
L =10
L =20
L =30
L =40
1.0 1.5
0.0
0.2
0.4
0.6
0.8
1.0
L
1.0 1.5
0.0
0.2
0.4
0.6
0.8
1.0
Playground

52. MACHINE LEARNING PHASES OF MATTER
Scientific paper link
Playground
??
??
IMAGE/Data ??

53. MACHINE LEARNING PHASES OF MATTER
Scientific paper link
Playground
??
??
IMAGE/Data ??
1.0 1.5 2.0 2.5 3.0 3.5
Temperature
UNSUPERVISED CLUSTERING
PCA, t-SNE, k-Means

54. MACHINE LEARNING FOR PHYSICS - REALLY ONLY STARTED ~2016
SHOWCASING THE BEST OF JAPAN’S PREMIER RESEARCH ORGANIZATION • www.riken.jp/en/research/rikenresearch
WINTER 2018
FLY BY THE SEAT OF THEIR PANTS
These insects find each other
via fecal pheromones
DREAM GENES
Two DNA markers control
our most vivid sleep state
SNAP FREEZE
Quick cooling reveals
new superconductors
M
ACHINE
LEARNING
Japan
is
poised
to
reap
the
benefits
ofcleverartificial
intelligence
algorithm
s
CONCLUSIONS AND OUTLOOK
▸ ML is applied to several branches of
physics: particle, statistical, astro…
▸ Physics is a complex problem with
plenty of opportunities for Deep
Learning algorithms
▸ Experimental and Computational
physics provide a very large amount
of data (sensors or simulations)
▸ Lead top500 supercomputers have
demonstrated amazing ML
capabilities and applications
▸ weather, earthquakes, etc…

55. EXTRA MATERIAL

56. Nature 558, 91-94 (2018)
Art by Bart-W. van Lith

57. Nature 558, 91-94 (2018)
Art by Bart-W. van Lith

58. MASSIVELY PARALLEL SUPERCOMPUTERS FOR IN-SILICO PHYSICS
LATTICE QUANTUM FIELD THEORY - OVERVIEW
‣DISCRETIZE SPACE AND TIME
‣ LATTICE SPACING a
‣ LATTICE SIZE L
‣KEEP ALL D.O.F. OF THE THEORY
‣ NOT A MODEL!
‣ NO SIMPLIFICATIONS
‣AMENABLE TO NUMERICAL METHODS
‣ MONTE CARLO SAMPLING
‣ USE SUPERCOMPUTERS
‣PRECISELY QUANTIFIABLE AND IMPROVABLE
UNCERTAINTIES
‣ SYSTEMATIC: L➝∞, a➝0
‣ STATISTICAL: √N

59. THE MATH BEHIND SIMULATIONS
LATTICE QUANTUM FIELD THEORY - MATHEMATICS
LQCD =
1
4
F2
+ ¯(i /D + m)
MICROSCOPIC THEORY
OF FIELDS
hOi =
1
Z
Z
D D ¯ DU e S[ ¯, ,U]
O
QUANTUM FEYNMAN
PATH INTEGRAL
+O
✓
1
p
N
◆
⇡
1
N
NX
i=1
O [Ui] IMPORTANCE SAMPLING
{U1, U2, U3, . . . , UN }
MARKOV CHAIN MONTE CARLO
DISCRETIZE
ψ: quark field
U: gauge field
Physical
observable
Makes integral finite dimens.
Estimator
Sampling

60. THE MATH BEHIND SIMULATIONS
LATTICE QUANTUM FIELD THEORY - MATHEMATICS
LQCD =
1
4
F2
+ ¯(i /D + m)
MICROSCOPIC THEORY
OF FIELDS
hOi =
1
Z
Z
D D ¯ DU e S[ ¯, ,U]
O
QUANTUM FEYNMAN
PATH INTEGRAL
+O
✓
1
p
N
◆
⇡
1
N
NX
i=1
O [Ui] IMPORTANCE SAMPLING
{U1, U2, U3, . . . , UN }
MARKOV CHAIN MONTE CARLO
DISCRETIZE
ψ: quark field
U: gauge field
Physical
observable
Makes integral finite dimens.
Estimator
Sampling

61. THE MATH BEHIND SIMULATIONS
LATTICE QUANTUM FIELD THEORY - MATHEMATICS
LQCD =
1
4
F2
+ ¯(i /D + m)
MICROSCOPIC THEORY
OF FIELDS
hOi =
1
Z
Z
D D ¯ DU e S[ ¯, ,U]
O
QUANTUM FEYNMAN
PATH INTEGRAL
+O
✓
1
p
N
◆
⇡
1
N
NX
i=1
O [Ui] IMPORTANCE SAMPLING
{U1, U2, U3, . . . , UN }
MARKOV CHAIN MONTE CARLO
DISCRETIZE
move in
configuration
space with prob.
ψ: quark field
U: gauge field
Physical
observable
Makes integral finite dimens.
Estimator
Sampling

62. THE MATH BEHIND SIMULATIONS
LATTICE QUANTUM FIELD THEORY - MATHEMATICS
LQCD =
1
4
F2
+ ¯(i /D + m)
MICROSCOPIC THEORY
OF FIELDS
hOi =
1
Z
Z
D D ¯ DU e S[ ¯, ,U]
O
QUANTUM FEYNMAN
PATH INTEGRAL
+O
✓
1
p
N
◆
⇡
1
N
NX
i=1
O [Ui] IMPORTANCE SAMPLING
{U1, U2, U3, . . . , UN }
MARKOV CHAIN MONTE CARLO
DISCRETIZE
move in
configuration
space with prob.
statistical error
ψ: quark field
U: gauge field
Physical
observable
Makes integral finite dimens.
Estimator
Sampling

63. NEW DIRECTIONS IN LATTICE QUANTUM FIELD THEORY WITH ML
MACHINE LEARNING IS INSPIRING NEW ALGORITHMS
▸ it is costly to generate
configuration with
MCMC in certain
regimes
▸ accelerate sampling with
generative models
▸ examples: RBMs,
normalizing-flow models,
GANs, self-learning
▸ unsupervised algorithms
can be used to detect
phase transitions in
materials
▸ reconstruct microscopic
parameters from
macroscopic observations
▸ spectral inference can be
used to find eigenfunctions
of quantum systems
GENERATION MEASURE

64. NEW DIRECTIONS IN LATTICE QUANTUM FIELD THEORY WITH ML
MACHINE LEARNING IS INSPIRING NEW ALGORITHMS
▸ it is costly to generate
configuration with
MCMC in certain
regimes
▸ accelerate sampling with
generative models
▸ examples: RBMs,
normalizing-flow models,
GANs, self-learning
▸ unsupervised algorithms
can be used to detect
phase transitions in
materials
▸ reconstruct microscopic
parameters from
macroscopic observations
▸ spectral inference can be
used to find eigenfunctions
of quantum systems
GENERATION MEASURE

65. NEW DIRECTIONS IN LATTICE QUANTUM FIELD THEORY WITH ML
MACHINE LEARNING IS INSPIRING NEW ALGORITHMS
▸ it is costly to generate
configuration with
MCMC in certain
regimes
▸ accelerate sampling with
generative models
▸ examples: RBMs,
normalizing-flow models,
GANs, self-learning
▸ unsupervised algorithms
can be used to detect
phase transitions in
materials
▸ reconstruct microscopic
parameters from
macroscopic observations
▸ spectral inference can be
used to find eigenfunctions
of quantum systems
GENERATION MEASURE
1. ORGANIZING (AI + LATTICE) WORKSHOPS
2. PAPERS IN PREPARATION ON USE CASES
FOR GAN AND PHASE TRANSITIONS
3. SUPERCOMPUTERS ARE AI-READY

66. GALACTIC ROTATION VELOCITY
GRAVITATIONAL LENSING
Rubin and Ford (1970)
Subaru Telescope
3D MapDark Matter