This is a Collaboration with the Imperial College Physics Society (Physoc)
- Task: Accurate prediction of X-ray pulse properties from a free-electron laser using Machine learning
- Fields: Lasers, optics, Solid State Physics
- Time and Date: 11:00-13:00, Saturday 4 December 2021
- Location: Lecture Theatre 2, Blackett Laboratory, South Kensington Campus and on ICDSS Teams Channel
- Presenter: Karim (4th Year Physics )
Details
The dataset is available from
https://github.com/alvarosg/DataLCLS2017
The work would consist mainly of extracting the right data from it, preprocessing it and then trying to make predictions. In the original paper about it, they used Linear, Quadratic, ANN and SVR to make predictions of different quantities. In principle, they can just follow that procedure, or try out other things. It is a really good dataset to learn the general skills of the trade, from data extraction to processing to supervised learning. From the physics point of view, we are basically looking at an X-Ray Free Electron Laser, specifically the LCLS in Stanford from where this data was collected (https://lcls.slac.stanford.edu/).
These are cool machines, because they can basically create laser pulses in the x-ray range of the spectrum. The issue is though, that they are absolutely massive (500m for LCLS, EU-XFEL in Hamburg clocks in at 3.4km), so we want to make the most out of these machines. We therefore want a high pulse rate, but that creates a problem. The spectrum of an individual pulse can vary greatly due to initial conditions, so we need to either measure or predict it. But for faster pulse rates, measuring it becomes unfeasible, requiring prediction.
An excellent paper on this by a previous researcher at imperial can be found here:
https://www.nature.com/articles/ncomms15461
(the data also stems from here, and this paper is based on the exact data that we would provide, paper results can be used as a benchmark).