What is a particle accelerator?
Did you know that there are more than 30,000 particle accelerators in operation around the world? Although they are mostly known to be used for research in physics, they have a variety of uses. Beams of charge particles can be used to treat cancer patients, to clean contaminated water and flu air, for sterilization of medical equipment and food irradiation, among other things. But what actually is a particle accelerator?
Particles can refer to charged atoms or charged sub-atomic particles (protons, electrons, etc.), carrying either positive or negative charge. While accelerators are machines built to literally accelerate any of these particles to speeds close to light-speed.
Magnetic fields are used for both, acceleration and aiming of the particle beam. When particle accelerators are built straight, they are called lineacs, and when they’re built as circles, they are called cyclotrons or rings. Particularly, circular accelerators can be used for collisions, in which two beams are directed in opposite directions and collide at detectors point.
In order to create and direct the beam(s), different parameters need to be adjusted, the energy of the beam, the number of particles, and the frequency of changing magnetic fields are just a few. They are usually continuous and involve complicated formulas, with an ever-higher demand of precision this task becomes harder for humans to solve. Thus, machine learning opens a possibility to automatize at least some tasks.
What is machine learning?
Machine learning is an application from artificial intelligence (AI), in which an algorithm performance for a task is optimized repeatedly "feeding" it data. The most common method used for modeling particle accelerators are neural networks. Neural networks do not need human intervention after the code has been written, it learns from their own mistakes. On a simple way, these networks contain functions (these are the so-called nodes or neurons) with weighted connections between them. A set of neurons form a layer, which process the data and feeds it to another layer. After it has been created, the network is “trained” to obtain desired results. The structure can be seen as a black box, the programmer gives the program training data with the corresponding correct result, after the program "learns" from this data, it is then ready to make predictions. The bigger the training data the more accurate the program becomes.
Implementig Machine Learning to Particle Accelerators
Because of the highly complex nature of accelerators, as demands require for better performance and stability, it becomes harder to adjust the settings using traditional control techniques (human operator or mathematical optimization) to satisfy the demands. One viable option is the use of machine learning and neural networks. The use of AI in particle accelerators is not new, it started during early 1990s, but it is the recent improvements in the technology and knowledge for its use what has make it resurface. These techniques can be used to automatically adjust the settings of accelerators when changing between operations, predict when a malfunction is about to happen, or make up for missing data. Some examples of how this tool has been used to automate operations on particle accelerators are described in the next session.
Examples of previous works
NEURAL NETWORK TO PREDICT BEAM PARAMETERS AND INPUT SETTINGS
They attempted to fine tune different devices at the beginning of the accelerator to obtain some desired parameters at another point, paying special attention on solving some inner forces between particles in the beam. The research used initially a neural network in which they would give the settings of the magnetic fields, and the program would predict the parameters for the beam at the entrance of an undulator (series of static magnets). The results were so promising, that they furthered the study for an inverted process (given the desired parameters it would change the settings). The neural network architecture varied in each stage, but both studies used reinforced learning, allowing the algorithm to learn and interact with the model. The model gave the operator the capacity to obtain a precise beam without the need to fine-tune by himself/herself each setting at the entrance of the particle accelerator.
NEURAL NETWORK TO ADJUST FREQUENCY FOR A COOLING SYSTEM
The frequency of resonance (the frequency for which the amplitude of the wave is highest) for a certain particle accelerator was controlled purely with a cooling water system. The team proposed an initial neural network model for the radio frequency of the system magnetic field apparatus, that allows a faster response to changes in temperature during pulsed operation (intermittent). They were able to create a model that predicts the resonant frequency of the RFQ under changes in the cooling system that performed sufficiently well for use in a daily control routine. They used a simple feed-forward architecture for the neural network, it was changed and refined multiple times. This method can be implemented to eliminate the need for a human operator to control such parameters, enabling the users to focus on bettering other areas, and could even help maintain a more stable temperature.
