A European network dedicated to AI for accelerator sciences and technologies

Adnan Ghribi

CNRS / GANIL

October 16, 2024

Outline

  1. Introduction
    Context, Purpose, State of the art, some history

  2. The project
    Scope, consortium, connections

  3. Excellence
    Objectives, ambitions & methodology

  4. Impacts
    Outcomes & synergies

  5. Implementation
    Concortium & organisation

  6. What happens next

Introduction

Context

What do we call AI (Artificial Intelligence) ?

That, you already know ! Right ?

Context

And how can AI help ?

Introduction

  • operation and reliability ;
  • Detecting, preventing anomalies ;
  • Optimising beam time ;
  • Frugal complex physics simulation ;
  • Improved twin models.
  • …

State of the art (1/2)

What people have been doing …

Introduction

Despite all the success stories, there are serious locks that prevent making global impact in the community !

State of the art (2/2)

Introduction

… and who these people are

Connection between a sample of 375 publications in AI and accelerators1.

Some history

Introduction

[March 2023]
First meetings indico
Several working groups

[July 2023]
Workshop at CERN ;
35 participants indico

[November 2023]
Workshop in Paris.
70 participants indico

We built working groups concentrating on ML accelerators related technical issues :

  • Optimization
  • Anomalies/Classification
  • Data generation

Finally, we tried to solve the puzzle …

A Route toward Sustainable Data Generation in Accelerator Science” ; doi:10.5281/zenodo.10685243

The consortium

The project

A network, a community
14 countries, 30 institutes,15 RI,7 ESFRI RI, 5 companies, many more people.

The consortium

The project

A network, a community, that extends beyond our walls

External observers from other countries and international organisations.

The consortium

The project

A network, a community, that extends beyond our walls and keeps extending beyond the frame of our projects

Direct connections and transverse contributions to several projects.

The consortium

Purpose

  • Organises future AI centred projects of the accelerator community in a synergetic way ;
  • Clusters activities, organises exchanges, meetings and workshops ;
  • Give feedback to RI and decision makers (ex. particle physics strategy, future RI french roadmap).

The consortium

And at the national level \(\rightarrow\) M4CAST

  • GENCI + CRIANN computing (under-used in 2024)
  • CCIN2P3 dedicated storage
  • Few PHd students
  • AISSAI1
  • ANR FUTURO
  • IN2P3 transverse project in preparation

Next M4CAST annual meeting on Nov 6th 2024

Open meeting but please register on indico if you want to come

Real life examples

The tip of the iceberg

The tip of the iceberg

The tip of the iceberg

ECHO STATE NETWORK FOR DYNAMIC APERTURE PREDICTION

Credit : Barbara Dalena

  • Blobby and light RNN type reservoir computing1 ;
  • Very fast dynamic aperture estimation (HL-LHC)2 ;
  • Could be used to predict indicators of chaos in large, complex future machines.

5e4 turns

Tracking –> 2h vs training+hyper-parameters search+predictions ~3 min ;

ML based Differentiable Beam Dynamics Simulation (Cheetah)1

Credits : Annika Eichler ; Jan Kaiser

  • Python based, Pytorch integration, Jaxed ;
  • Autodiff beam dynamic gradient computing (ex. fast space charge)
  • Seemless integration of NN surrogates, BO optimisation prior

Other toughs

  • GNN representation
  • CNN representation
  • RNN prediction based
  • Physics constrained
  • Shared latent space
  • And the Holy Graal

`

particle-particle interaction

`

beam-line elements correlations

`

Tracking speed-up

`

both NN structure and loss constrained

`

heterogeneous pre-trained models / heterogeneous data / single model integration

`

An accelerator physics foundation model

Conclusion

  • We need efficient/precise, fast/frugal simulation tools for future accelerators ;
  • ML based methods intertwines naturally with physics simulations ;
  • There is a lot of potential but there is lots of work to do ;
    • and for that, we need synergy !

Thank you

Questions ?

Atelier Calcul GDR SCIPAC, Orsay, October 16, 2024