Master's thesis: explainable machine learning for quench detection

During my master's thesis I studied the behavior of a family of magnets called High Order Correctors. These magnets have been mounted on the Large Hadron Collider and they are utilized in the relative proximity of the interaction regions to make sure that the beam that is going to have to pass through the machinery doesn't present any kind of unwanted oscillations on different magnetic field harmonics.

High Order Correctors, like any other superconducting magnet, need to be operated within certain operational parameters. This means that, in order to maintain the electrical resistance at 0, we need to make sure that:

If the temperature raises too much (due to mechanical stress or other) or the applied magnetic field is too strong due to the presence of neighbouring magnets within the assembly, then the magnet could quench.

Quench is a phenomenon that happens whenever a magnet transitions from the superconducting state (resistance 0) to the conducting state (resistance >0). Whenever that happens we might have an irreversible buildup of heat within the magnet that could catastrophically damage it and the machine.
In order to avoid thousands upon thousands of euros of damage to the machine or its devices, a failsafe mechanism was designed to make sure that, whenever an increase in the resistive voltage is being picked up, energy is completely cut from the magnet and the beam itself is redirected to a lead target to be dumped safely. This avoids any damage to the machine and the infrastructure and it gives time to the magnets to cool off.

Our idea was to see whether we were able to perform a more accurate analysis by using machine learning models. Can we use machine learning to accurately predict what kind of quench we are in front in order to reduce the amount of beams dumped? Can we develop machine learning models capable of helping in the design of future magnets?

At the end of our research, we were able to make an analysis on quadrupolar data for High Order Correctors and we found a strong correlation between the information obtained via post-mortem analysis on the harmonic decomposition of the magnetic field and the presence and location of the quench within the magnet's assembly.
We used simple models like decision trees and random forests, we then created our own "collage model" that was able to perform slightly better than the default scikit-learn models.

The research we conducted was further developed by PhD students from Naples University that developed models based on isolation forests and Autoencoder/Decoder for anomaly detection which helped identifying the quench phenomenon, isolating it from noise, using data coming from voltage taps placed within the magnet assembly (a different kind of data from the one that I worked with) and unlocking the possibility to recognize this problem with high performance in a runtime environment.
Currently the biggest question remains, of course, unsolved. How can we extend this to a generic magnet? Our studies proved to be useful only for High Order Correctors of the quadrupolar kind, that is due to the fact that the distribution of the odd cases was extremely skewed (over 200 cases for quadrupoles, less than 50 for hexa, octa, deca and dodecapoles). Furthermore the values that the models deployed learned are not useful for different kinds of magnets, which reach different magnetic field levels and therefore have different harmonic decomposition of the residual magnetic field after the Quench Protection System kicks in.