1 of 18

Panel Discussion

First COMETA workshop on artificial intelligence for multi-boson physics, Nihkef, Amsterdam, NL

2 of 18

Today’s session

2

3 of 18

Today’s sessions - Objectives

  • Discuss open problems & brainstorm on best AI/classical tools
  • Focus on two/three main areas:
    • Flavour tagging
    • Jet Substructure
    • (Polarisation) [linked to other COMETA activities]
  • How can COMETA effectively help improve state-of-the-art ?
    • What tools/datasets do we need to foster collaborations ?
    • In which areas can we improve/complement existing effort ?

​

  • Practicalities - concerning size of datasets, speed of inference, calibration etc.
  • Performance - ideas for improvement, new models, links with theory, conditioning, bias etc.
  • Physics - effect/dependence of these algorithms on results, which information is learned, which measurements might help etc.

3

4 of 18

FLAVOUR TAGGING

4

5 of 18

Advances in experiment and theory

5

Talk by Francesco yesterday ; talk by Simone yesterday

6 of 18

Advances in experiment

6

Talk by Francesco yesterday

Differentiable programming [2310.12804]

Mask Former [2312.12272]

Foundation models [2405.12972]

7 of 18

Trigger developments in ATLAS & CMS

7

Talk by Francesco yesterday

8 of 18

Wealth of activity … ready to be expanded

  • Flavour- and jet-tagging showcase impact of ML for LHC physics
  • Progress in theory and experiment (offline and trigger)

​

  • Many developments done within experimental collaborations (with data), but many results showcasing capabilities of new algorithms
  • Can we define and share datasets to foster further collaborations?
    • Is Monte Carlo good enough / easy enough to obtain ?
    • Or, is data essential ? (And how to obtain such datasets ?)
    • What can we do within experimental collaborations (or outside) ?

8

9 of 18

WHAT CAN WE DO WITHIN COMETA?

9

  • Practicalities - concerning size of datasets, speed of inference, calibration etc.
  • Performance - ideas for improvement, new models, links with theory, conditioning, bias etc.
  • Physics - effect/dependence of these algorithms on results, which information is learned, which measurements might help etc.

10 of 18

JET SUBSTRUCTURE

10

11 of 18

Very active area of development

11

Talk by Mario yesterday

[2004.08262]

Algorithms comparison CMS

12 of 18

Wealth of activity … ready to be expanded

  • Jet physics extremely important at LHC
  • ML is ubiquitous for reconstruction and calibration
    • Will require lots of ingenuity … and external expertise

​

  • Many developments done within experimental collaborations (with data), but many results showcasing capabilities of new algorithms
  • Can we define and share datasets to foster further collaborations?
    • Is Monte Carlo good enough / easy enough to obtain ?
    • Or, is data essential ? (And how to obtain such datasets ?)
    • What can we do within experimental collaborations (or outside) ?

12

13 of 18

WHAT CAN WE DO WITHIN COMETA?

13

  • Practicalities - concerning size of datasets, speed of inference, calibration etc.
  • Performance - ideas for improvement, new models, links with theory, conditioning, bias etc.
  • Physics - effect/dependence of these algorithms on results, which information is learned, which measurements might help etc.

14 of 18

POLARISATION

14

15 of 18

Polarisation(-aware) taggers for hadronic vector bosons

COMETA workshop on vector-boson polarisations [https://indico.cern.ch/e/1371888/]

  • Amplitude-assisted tagging of longitudinally polarised bosons using wide neural networks [2306.07726]: use amplitude to extract theory parameters / pseudo observables from data (talk by M. Pellen)
  • Experimental perspectives for polarisation tagging (talk by A. Giannini)
    • Time to tackle hadronic decays of V bosons…

15

16 of 18

16

17 of 18

Wealth of activity … ready to be expanded

  • Can we define and share datasets to foster further collaborations?
    • Is Monte Carlo good enough / easy enough to obtain ?
    • Or, is data essential ? (And how to obtain such datasets ?)
    • What can we do within experimental collaborations (or outside) ?

17

18 of 18

WHAT CAN WE DO WITHIN COMETA

18

  • Practicalities - concerning size of datasets, speed of inference, calibration etc.
  • Performance - ideas for improvement, new models, links with theory, conditioning, bias etc.
  • Physics - effect/dependence of these algorithms on results, which information is learned, which measurements might help etc.