# Helmholtz AI for accelerator physics - an introduction and a showcase
Intelligent Process Control Seminar, March 9, 2021
# Questions and Notes
Use this hackmd pad for notes or questions during the talk! **Contributions are anomynous!**
Q: clustering algorithm requires to set the number of clusters?
- (H) yes, kmeans requires this; inline with the task of anomaly detection (healthy and damaged state), Helene decided for 2 clusters;
- (H) can also cluster to more than 2 states -> obtained from looking at the data or models/concepts
- provided data in different operation modes -> data was somewhat diverse, but still the data available is from two different lasers
- (H) new data expected, will be curious to see development
Comment: data taken from actual XFEL during operation. Not feasible to generate clean data.
- unclear in how many different states the laser is in
- number of states could be defined from first principles/theory
- even with same configurations, the laser might not reproduce the data at any point of time
- on top: laser drifts slowly through time -> goal: observe feature space to see when laser quality drops
Q: how do you ensure that the data is healthy? You may be learning from a broken laser already.
- during Nov data taking until shutdown -> decided that this is healthy laser
- broken state obtained from frequ. components in "inloop observables"
- future: model might tell us?
Q: how sensitive are the algorithms on time synchronous dynamics? (fixed sampling rate/time)
- do you mean towards online detection?
- with a dynamic system, with a fixed sampling time ??
- data taken from "dukes" devices and cleaned inbetween -> trimmed to .1 seconds chunks
- knowing the time scale is important
Comment: Voucher system works quite well. Tricky to set up goals in the beginning.
Comment: Important to have good data available. Also to shape expectations. Worked well!
Comment: common language was a challenge. Amazing to see prototype come around so quickly and now we have to validate the model and results.
## Share with us something that you liked or that you learned :heavy_check_mark:
## Share with us something that you didn't like or that you want us improved :exclamation: