1. LHC beam mode classification
Dr Ing Gianluca Valentino
Department of Communications and Computer Engineering
University of Malta
Data Science Research Group

2. Outline Introduction – LHC Machine Cycle Features used
LSTM model training
Results
Conclusions

3. Introduction – LHC Machine Cycle
Adjust
Injection
Ramp
Flat Top
Stable Beams
Squeeze

4. Problem formulation
We want to build a classifier which can predict transitions amongst the following four beam modes using only beam loss data:
Transition between injection to ramp
Transition between ramp and flat top
Transition between flat top and squeeze
Transition between squeeze and collisions

5. Dataset generation
There are ~3600 Beam Loss Monitors (BLM) around the LHC to measure local beam losses.
They provide 1 Hz data at various running sums (RS09 = 1.31 s chosen, similar to what is normally used in multi-turn loss analysis.
Data obtained from Timber (CERN logging service), from 168 pp physics fills.
+/- 50 seconds around change in beam mode.
Considered only losses from the 42 BLMs at IR7 collimators.
Therefore each training sample has dimension 100 x 42.
BLM Ionization Chambers in LHC

6. Start Squeeze
Start Flat Top
Some examples of the multivariate time-series BLM data
Start Ramp
Start Adjust

7. Dataset generation Feature scaling: Summary of classes:
Beam mode transition
# samples in dataset
Start of ramp
168
Start of flat top
166
Start of squeeze
131
Start of adjust
151
Total
616
Feature scaling:
Each multivariate time series was normalized by the BLM signal at the TCP.C6L7.B1 (primary) collimator - where we generally expect highest losses.

8. Recurrent Neural Network
Motivation: Not all problems can be solved with a neural network structure having a fixed number of inputs and outputs.
Measurement data is often sequential (time-series)
In practice we can have different input/output scenarios

9. Recurrent Neural Network
Output Sequence
Output Yt
Yt-1 Yt
Yt+1 wy wy wh wh … … = ht
ht-1 ht
ht+1 wx wx Xt
Xt-1 Xt
Xt+1
Input
Input Sequence

10. Problem of long-term dependencies
Consider the difference between:
“The clouds are in the sky”.
“I lived in France for three years while I was working for a software development company. I can speak fluent French”.
Classical RNNs are not capable of learning these long-term dependencies -> we need LSTMs.

11. Long Short Term Memory (LSTM)
Instead of a single neural network, there are four networks.

12. LSTM cell state
The cell state is the horizontal line running across the top
Information can flow along it unchanged or with minor modifications.

13. “Forget gate” layer
Decides which values of cell to reset

14. “Input gate” layer
Sigmoid layer: decides which values of cell to write to
Tanh layer: creates vector of new candidate values to write to cell

15. Update cell state
The LSTM applies the decisions to the memory cell:

16. “Output gate” layer
A sigmoid layer decides which values of cell to output.

17. ML training A RNN-LSTM model was trained to predict the output class.
The output of the LSTM was forwarded to a single Dense layer of size 4, each with a softmax activation function
One-hot encoding was used to represent the output (e.g. class #2 -> [0,0,1,0])
Train/test ratio used: 80% / 20%.
Implementation: Python and Keras.

18. ML training
Cross-validation was done to determine the best parameters for the following:
Number of LSTM neurons: [8, 16, 32, 64, 128, 256] -> 32 picked as best
Optimizer: [Adam, SGD, RMSprop, Adadelta] -> Adam picked as best
Dropout: [0.1, 0.2, 0.3, 0.4] -> 0.2 picked as best

19. Results
Accuracy and loss on training and testing sets

20. Results Classification Report from Scikit-Learn: Beam mode transition
precision
recall
f1-score
Start of ramp
1.00
0.94
0.97
Start of flat top
0.83
0.89
Start of squeeze
0.75
0.86
0.80
Start of adjust
0.72
0.84
0.78

21. Conclusions
Demonstrated applicability of LSTMs in classifying transitions in beam mode
First time (I believe) RNNs used on LHC data
Achieved accuracy on unseen data of ~87%
Future work: try also using orbit data from BPMs