10th MT ARD ST3 - pre-meeting Machine Learning Workshop

Wednesday 7 Sept 2022, 08:30 → 12:00 Europe/Berlin

NEW 15 1'427 (Humboldt University of Berlin)

Andrea Santamaria Garcia (KIT)

Data-driven methods such as machine learning (ML) can accurately reproduce the behavior of a mathematical model at a substantially reduced computational cost, given enough data. Not only that, but they have the potential to outperform classical simulations. In accelerator physics, the reduction in computational cost through ML models opens the door to real-time deployment of online systems not only for beam control during complex phenomena but also for online accelerator optimization and prediction.

The accelerator physics community is already actively applying a variety of ML methods to many different problems. This workshop offers a very applied introduction to ML for accelerators, including an overview of algorithms and tools, technical lectures, hands-on tutorials, and real-life applications.

As this workshop is part of the annual MT-ARD-ST3 meeting: please don't forget to register to the main event and pay the fee!

-->  https://indico.desy.de/event/33584/

Target audience:

• It is meant for people with no previous experience in machine learning
• Previous experience with Python and Jupyter notebooks is required!

Logistics:

• We will release the code for the workshop on GitHub before the event
• There will be computers available to run the code, but you can also set it up beforehand in your own laptop

Material:

• https://ansantam.github.io/2022-MT-ARD-ST3-ML-workshop/
• https://github.com/ansantam/2022-MT-ARD-ST3-ML-workshop

08:30 → 08:45 Welcome and introduction to machine learning in accelerator physics

Speaker: Andrea Santamaria Garcia (KIT)

0-welcome.pdf

08:45 → 09:05 Introduction to artificial neural networks

Speaker: Andrea Santamaria Garcia (KIT)

1-neural-networks.pdf

09:05 → 09:45 Coding example: build your own neural network

We will fit non-linear functions with neural networks in PyTorch and understand the role that the different parameters of the model play in the quality of the fit

09:45 → 10:00 Special topic: introduction to Bayesian optimization

Speaker: Chenran Xu (KIT)

2-bayesian-optimization.pdf

10:00 → 10:30 Coding example: optimize unknown functions with Bayesian optimization

We will implement all the basic components of Bayesian optimization (BO), and see how to use BO for some sample 1D and 2D functions

10:30 → 10:45 Coffee break

10:45 → 11:00 Application of Bayesian optimization to improve injection efficiency at KARA demo

Speaker: Chenran Xu (KIT)

3-bo-kara-demo.pdf

11:00 → 12:00 Special topic: introduction to reinforcement learning & ARES demo

Speakers: Jan Kaiser (DESY), Oliver Stein (MSK (Strahlkontrollen))

3-reinforcement-learning.pdf