The Large Hadron Collider (LHC) is set to improve its cleaning efficiency with the use of the innovative technique known as crystal collimation. This technique utilizes bent crystals to deflect halo particles into dedicated collimator absorbers, which helps to increase the betatron cleaning efficiency with high-intensity ion beams. To achieve optimal channeling conditions, the crystals must be kept in perfect alignment with respect to the circulating beam envelope. This can be particularly challenging given the small angular acceptance.

To overcome this challenge, we have developed a new machine learning algorithm which is a step forward in the field of crystal collimation, by using a 1D -convolutional neural network. The machine learning model is expected to optimize the efficiency of the crystal collimation setup in future applications of this beam collimation technology.

This original technique utilizes the unique property of materials with highly ordered atomic structures to capture charged particles with suitable impact conditions in the potential well generated by neighboring crystalline planes, a process called crystal channeling. Therefore, bent crystals are used to efficiently steer beam halo particles by forcing them to follow the curvature of the crystal itself.

The CEM MRO team's efforts were put to the test in a recent experiment using Pb Ions beam at a record energy of 6.8 Z TeV, where the algorithm demonstrated its reliability.

Dataset

The optimal channeling orientation is identified using Beam Loss Monitors (BLMs) while the crystal is slowly rotated (angular scan). Figure 2 shows that the channeling signature is composed of three main patterns: (1) amorphous plateaus where the orientation is so far away from optimal channeling that the crystal behaves like a standard collimator; (2) volume reflection plateau where particles bounce from the crystalline planes instead of being channeled; (3) and the channeling well, a minimum in the loss pattern observed at the crystal location due to the beam particle being channeled by the crystal. The dataset used in this work consists of 1332 sets of 1 Hz BLM signals gathered during machine development studies with proton and Pb ion beams from 2015 to 2022. The segmented BLM signals have been distributed into a main dataset used for training the CNN and a validation set used for testing the model on unseen data. Both datasets have been divided in three classes. In particular, the distribution has been performed in the following way:

1. Main Dataset:

• “Channeling Well” Class: 265 signals;
• “No Well” Class: 444 signals;
• “Partial Well” Class: 394 signals.

1. Validation Set:

• “Channeling Well” Class: 54 signals;
• “No Well” Class: 124 signals;
• “Partial Well” Class: 51 signals.

A signal belonging to class “Channeling Well” (Figure 2) is a BLM signal that shows the channeling signature described above. A signal belonging to class “No Well” (Figure 3) is a signal in which the channeling pattern cannot be identified. A signal belonging to class “Partial Well” (Figure 4) is a signal in which can be identified the pattern of channeling but not the volume reflection.

Machine Learning Model

The machine learning model trained to classify crystal collimators channeling has been a 1D convolutional neural network (1D-CNN), that are commonly used on sequential or time series data. CNNs have the ability to learn complex objects and patterns and low-cost hardware implementations due to the simple and compact configuration that perform only 1D convolutions (scalar multiplications and additions). The structure of the CNN listed in Table 1, originally proposed in [2], was developed with the use of the deep learning library Keras [3] with TensorFlow [4] for the back end and it has been adapted to the problem under study. Before feeding the data into the first CNN layer a Z-Score normalization at each signal is applied, such that they have the properties of a standard normal distribution with mean µ = 0 and standard deviation σ = 1 [5]. The developed model consists of a 1D convolutional layer followed by a batch normalization layer, a rectified linear unit activation function and a dropout layer (with a 0.15 frequency rate) adopted to reduce overfitting. The aforementioned structure is repeated three times and is closed by a 1D global average pooling layer and a dense layer with three neurons with a softmax activation function. The choice of the latter allows to output three different probabilities, precisely the probabilities that the time series analyzed shows a pattern compatible with Figure 2, Figure 3 or Figure 4.

The 1D-CNN performance has been evaluated with the use of the precision metric , i.e. the ratio between true positives and the sum of true positives and false positives. Where true positives (TP) indicate the number of signals of the class that are correctly predicted by the algorithm and false positives (FP) indicate the number of signals not belonging to the class that are mistakenly classified as belonging to the class. The 1-D CNN achieved a weighted average precision of 91% (Figure 5) on unseen validation set BLM signals with proton and ion beams. This is a promising result, and indicates that convolutional neural networks can be applied to tackle the task of finding the optimal channeling.

Deployment

A continuous classification scheme has been engineered. The latter, uses a classification window determined with the use of the crystal’s bending angle and step size (Figure 6).

Where the classification window is the signal fed into the CNN, the bending angle is the angular bending of the crystal (defined as θ = Rl where l is the crystal length in the beam direction and R its bending radius) and the step size is the movement performed every second during the angular scan. Updating the data in this window continuously as the signal comes from the BLM signal located in proximity of the crystal it’s possible to obtain real time classification and by monitoring the probabilities in output from the CNN crystal channeling can be detected. An animation regarding this process is depicted in Figure 7.

As it can be seen from the animation in Figure 7 the algorithm first detects a partial well and right after it detects a channeling well with a probability over 90%. The ML model has been implemented in the software that controls the crystal goniometer and tested in real operation with Pb ion beams detecting for the first time ever crystal channeling with the use of machine learning (6).

This development has shown good results in terms of precision and demonstrated reliability in the application of machine learning in the context of automating crystal collimator alignment at the Large Hadron Collider (LHC). For this reason, it has been decided to implement the deep learning model in the operational crystal software framework.

Future work in this area will aim to optimize the continuous monitoring of losses while the crystal is kept in channeling. The goal of this task is to recognize if the optimal orientation is being lost, not only because the crystal is moving, but also due to changes in the beam dynamics. Being able to adapt and compensate for these changes is crucial to ensure stable performance of crystal collimation during operation.

References

[1] M. D’Andrea, G. Azzopardi, M. Di Castro, E. Matheson,D. Mirarchi, S. Redaelli, G. Ricci, G. Valentino. Prospects to apply machine learning to optimize the operation of the crystal collimation system at the LHC. IPAC, 2022
[2] H.I. Fawaz, G. Forestier, J. Weber, L. Idoumghar and P.A. Muller, ”Deep learning for time series classification: a review”, in Data Mining and Knowledge Discovery vol. 33, pp. 917–963, March 2019.
[3] F. Chollet et al., ”Keras”, https://github.com/fchollet/keras, 2015.
[4] M. Abadi et al., “Tensorflow: A system for large-scale machine learning”, in Proc. OSDI’16, Savannah, GA, USA, Nov. 2016.
[5] R. Vidiyala, “Normalization vs Standardiza-tion”, 2020. https://towardsdatascience.com/ normalization-vs-standardization-cb8fe15082eb.