5. Experiment is everything
〉Aristotle “proved” heavier objects
fall faster.
〉His proof was elaborate and well-
thought-of
〉It was undisputed until Galileo
〉Is refuted by a simple experiment
5
Image credit: http://www.liveyourmagic.com/
7. Inside (and around) a black hole. Quantum gravity.
7
Unsolved problems in Physics
Image credit: NASA
8. We made a golf ball with holes (empirical), but still can’t
calculate the airflow
8
Turbulence
Image credit: CD-adapco
9. History of Physics — Galilei, XVII
〉Disproved Aristotle’s theory of
gravity
〉Dropped balls of different weight
〉Took paper notes
9Image credit: heritage-history.com
10. History of Physics — LHCb, 2015
〉LHCb registers 10 mln. collisions
per second
〉Extensive computer system is used
for acquisition, filtering, storage and
analytics
10
11. 〉Results of interactions, including proton collisions, are fundamentally
random. The higher is energy, the more variants are possible.
〉Unfortunately, standard model (SM) explains all the available experiments
— except gravity. We continue trying — Beyond Standard model (BSM),
supersymmetry (SUSY).
11
Two words on state-of-the art in particle
physics
13. 〉The unit of data
〉Results from a collision of two proton bunches
〉In the beginning has raw detector readings
〉Has reconstructed information on the detected product particles —
angles, type, impulses
13
Event
14. 〉Formulate prediction. Usually, probability of a specific decay channel.
〉Test it.
〉Events we are looking for are very (orders of magnitude) rare.
14
Analysis — from data to article
15. 15
Analysis — data processing pipeline
Detector Triggers GRID
Event Index
107 events/s
Offline
analysis
103 events/s
Publication
16. 16
Analysis — data processing pipeline
Detector Triggers GRID
Event Index
107 events/s
Offline
analysis
103 events/s
ML
ML
Publication
18. 〉Remember Standard Model?
〉Use Monte-Carlo (MC) to generate samples of signal (interesting) and
background (not interesting) events.
〉Train classifiers.
〉Voilà! Aller à Opéra.
18
How to apply Machine Learning?
19. 〉Remember Standard Model?
〉Use Monte-Carlo (MC) to generate samples of signal (interesting) and
background (not interesting) events.
〉Train classifiers.
〉Voilà! Aller à Opéra.
〉Fail. Background is too diverse for generation. Combinatorial.
Computationally expensive, wasteful - we’ll throw away the majority
generated events. Plus some complicated stuff called “calibration”.
19
How to apply Machine Learning?
20. 20
How to test a hypothesis?
〉Define signal decay
〉Calculate expected event count in its mass window in background-only
assumption
〉Generate signal MC
〉Use the mass cut to select real background from sidebands
〉Fit the classifier on MC signal and real background
〉Apply the classifier to the target region, find the actual event count
〉If it’s inconsistent with background-only hypothesis, you have a discovery.