1 of 46

Estimation sensitivity for HEP

via Point Cloud Pooling with Graph Neural Networks

MOCA 2023 Jose M Munoz

11/05/2022

2 of 46

Estimation sensitivity for HEP

via Point Cloud Pooling with Graph Neural Networks

2

MOCA 2023 Jose M Munoz

3 of 46

¿Who am I?

3

4 of 46

4

Brief Introduction

1.

Context

State of the art

2.

ML and DM

4.

Learning Substructure

3

Our Proposal

Of the approach

Results

& Conclusions

5.

Future Outlook

Jose M Munoz, et al.

MOCA 2023 Jose M Munoz

5 of 46

Context

The why behind all of it

5

MOCA 2023 Jose M Munoz

6 of 46

The presence of DM: A deep motivation

6

MOCA 2023 Jose M Munoz

7 of 46

The presence of DM: A deep motivation

7

XENON Collaboration

MOCA 2023 Jose M Munoz

8 of 46

Parameter constraints

8

Betancur, el al. 2022

Marrodán, el al. 2015

MOCA 2023 Jose M Munoz

9 of 46

Complex Models Imply Complex Data

9

Chadha, 2021

MOCA 2023 Jose M Munoz

10 of 46

So… what we do?

10

Theory Parameters

Parton Level

Hadronization

Detector Effects

Observables

MC sample

Significance

MOCA 2023 Jose M Munoz

11 of 46

The Cut and Count Method

11

Aarrestad, et al. 2021

Cowen, 2017

MOCA 2023 Jose M Munoz

12 of 46

The Cut and Count Method

12

Aarrestad, et al. 2021

MOCA 2023 Jose M Munoz

13 of 46

Counting Experiments

13

MOCA 2023 Jose M Munoz

14 of 46

02. ML + DM

Previous Approaches

14

MOCA 2023 Jose M Munoz

15 of 46

Simply use ML to build an observable

15

Brehmer. 2020

Schwartz. 2022

MOCA 2023 Jose M Munoz

16 of 46

16

Abudinen, et al.

MOCA 2023 Jose M Munoz

17 of 46

17

MOCA 2023 Jose M Munoz

18 of 46

18

MOCA 2023 Jose M Munoz

19 of 46

19

MOCA 2023 Jose M Munoz

20 of 46

20

MOCA 2023 Jose M Munoz

21 of 46

21

22 of 46

03. GNN sense

Advocating for Pooling Graphs

22

MOCA 2023 Jose M Munoz

23 of 46

What we actually see in nature

23

MOCA 2023 Jose M Munoz

24 of 46

But what is really the log likelihood?

24

Cowan. 2022

MOCA 2023 Jose M Munoz

25 of 46

Problems of traditional ML

25

MOCA 2023 Jose M Munoz

26 of 46

One option is always padding

26

MOCA 2023 Jose M Munoz

27 of 46

One option is always padding

27

MOCA 2023 Jose M Munoz

28 of 46

One option is always padding

28

MOCA 2023 Jose M Munoz

29 of 46

One option is always padding

29

MOCA 2023 Jose M Munoz

30 of 46

The idea of Point Clouds

30

MOCA 2023 Jose M Munoz

31 of 46

The idea of Point Clouds

31

MOCA 2023 Jose M Munoz

32 of 46

GNN: Self learns

32

MOCA 2023 Jose M Munoz

33 of 46

A better estimation of LLR (More Direct)

Event based classifier

Ours

33

MOCA 2023 Jose M Munoz

34 of 46

04. Results

Toy Models and DM

34

MOCA 2023 Jose M Munoz

35 of 46

A toy Model

35

MOCA 2023 Jose M Munoz

Anzalone, et al. 2022

36 of 46

A toy Model

36

MOCA 2023 Jose M Munoz

37 of 46

Powerful Rejection rates

37

MOCA 2023 Jose M Munoz

38 of 46

Powerful Extrapolation Capacities

38

MOCA 2023 Jose M Munoz

39 of 46

Robust convergence

39

MOCA 2023 Jose M Munoz

40 of 46

05. Conclusions & Outlooks

Toy Models and DM

40

MOCA 2023 Jose M Munoz

41 of 46

Parametric extrapolation inference

41

Schwab, 2020

MOCA 2023 Jose M Munoz

42 of 46

Expand tests to more real DM models

42

MOCA 2023 Jose M Munoz

43 of 46

Thanks!

Questions?

43

MOCA 2023 Jose M Munoz

44 of 46

44

MOCA 2023 Jose M Munoz

45 of 46

45

MOCA 2023 Jose M Munoz

46 of 46

46

MOCA 2023 Jose M Munoz