Participants
Xiaodan Li, Rodrigo Gonzales-Gonzaga, Gian Luca Orlandi, Christopher Arrell, Rasmus Ischebeck, Thomas Schietinger, Florian Löhl, Franziska Frei, Andreas Adelmann, Daniela Kieselev, Nicole Hiller, Johannes Kirschner, Christoph Kittel, Lianglign Shi, Benedikt Hermann, Daniel Llorente, Cigdem Ozkan Loch, Simona Bettoni, Davide Reggiani, Jochem Snuverink, Hui Zhang
Introduction
Andreas Adelmann: classification of different approaches, history of machine learning and neural networks, building blocks of neural networks, topology of neural networks, learning (training): unsuipervised learning, reinforcement learning, supervised learning, examples
ideas for using ML for HIPA:
> reduce the number of interlocks by reducing losses
> model for spot size prediction on the SINQ target (Davide)
Use the "theano" ML framework
Collaboration with Auralee Edelen
ML for SwissFEL
Franziska, Nicole, Johannes:
> diagnostics for SwissFEL
> parameters that influence the properties of the X-ray pulses given to the users
> response of diagnostics on these parameters
> tuning and optimization of SwissFEL
> Bayesian optimization
> constraints: keep losses low
> implement existing Ocelot framework from DESY/SLAC
> expand into ML-based failure detection?
> introduction of Andreas Krause's group at ETHZ: "Decision making under uncertainty"
> Johannes' PhD: theory and algorithms for exploration-exploitation, e.g. optimize unknown function through noisy evaluations
Discussion
What happens if operators lose knowledge because they are supported by ML tools in routine operation?
–> virtual machine?
Connection to "conventional" feedback algorithms
Biology
Xiaodan Li
Problem: solve protein structure
Tools: SLS, SwissFEL, electron microscope
Drug development: cancer
Membrane proteins
Interaction between proteins
Applications of ML for biological problems:
> drug selectivity
> screening of potential drugs
Discussion
> Identify specific problems
> Literature search
> Common infrastructure
Application to Beam Position Monitors (DESY)
Liangliang Shi
> Wake fields in superconducting cavities
> Use neural network to predict orbit of beam
> Web site: deep learning for self-driving cars, hobby game with mini-drone
> G. Hinton, 'dynamic routing between capsules'
> Motivation: liberate accelerator physicists from tedious operation work to do more meaningful work
Photon diagnostics: application for photon diagnostics
Christopher Arrell
> ARAMIS: 30 independent devices to analyze photon beam properties
> ATHOS: situation will be even more disconnected
> Use ML to improve reliability of measurements
> for example, measurement of two wavelengths as well as temporal delay in double-pulse operation mode
> Accurate prediction of x-ray pulse properties from a free-electron laser using machine learning
> LCLS 2: use ML to link slow and fast diagnostics
> SwissFEL: improve accuracy of measurements
Plans
Nicole Hiller
Next meeting:
E-Mail list
Andreas Adelmann: write a document on pland for ML at PSI –> FOKO
There are minutes attached to this event.
Show them.