Application of sequence learning for predicting radiation damage of the CMS electromagnetic calorimeter

At the heart of modern physics lies the quest to understand the fundamental building blocks of the universe. To do this, scientists smash particles together at incredible speeds, creating conditions that existed just moments after the birth of the cosmos. One of the most powerful machines built for this purpose is the Large Hadron Collider, a massive ring buried beneath the border of France and Switzerland. Inside this machine, the Compact Muon Solenoid experiment acts as a giant, ultra-sensitive camera, snapping pictures of the debris from these high-speed collisions. A critical part of this camera is a detector made of thousands of heavy, clear crystals. When a particle hits one of these crystals, it releases a flash of blue light, much like a stone skipping across a pond creates ripples. By measuring the brightness of this light, physicists can calculate the energy of the particle that struck it. However, the environment inside the collider is incredibly harsh. The constant bombardment of radiation acts like a slow, invisible sandblaster on the crystals, gradually clouding their transparency. Over time, this damage makes the crystals less efficient at transmitting light, which threatens the accuracy of the entire experiment. To keep the detector working, scientists must constantly monitor how much the crystals have clouded and correct their measurements accordingly.

In a recent study, researchers turned to artificial intelligence to solve the problem of predicting how these crystals age. Instead of relying solely on manual checks or simple statistical guesses, the team applied a sophisticated type of machine learning known as a sequence-to-sequence model. This approach treats the history of a crystal's transparency as a story with a beginning, middle, and end. The computer is fed a long list of past measurements, showing how the crystal's clarity changed day by day as the collider ran and then stopped. The system learns to recognize patterns in this data, such as how the crystals darken rapidly during periods of intense collisions and how they slowly recover their clarity during quiet periods when the machine is not running. The goal was to teach the computer to read this history and then write the next chapter, predicting how the crystal would behave in the future.

The researchers tested this method using real data collected from the collider between 2016 and 2018. They focused on specific crystals within the detector, feeding the computer a window of past calibration values and asking it to forecast the transparency for the days ahead. The results were promising. The artificial intelligence model successfully tracked the rise and fall of the crystals' clarity, matching the actual measurements with a high degree of accuracy. For the crystals analyzed, the predictions showed a mean absolute percentage error of approximately 18% to 26%, while the absolute error remained within a manageable range. The model proved particularly good at short-term forecasts, capturing the immediate effects of radiation damage and the subsequent recovery. This suggests that such systems could become a reliable tool for the scientists running the experiment, allowing them to anticipate damage before it significantly impacts their data.

The study also took a closer look at previous attempts to solve this same problem using machine learning. The researchers discovered that an earlier study had reported a strange dip in performance at certain data intervals, which initially looked like a surprising scientific anomaly. Upon re-examining the code used in that earlier work, the team found a simple programming error that had caused the misleading results. Once this bug was fixed, the previous model's performance became much more consistent, though it still did not match the stability of the new sequence-to-sequence approach. The new model maintained a steady level of accuracy regardless of how much past data was used, showing a more robust ability to generalize and predict future behavior.

This work demonstrates that deep learning can effectively model the complex, time-dependent aging of detector materials. By accurately predicting how the crystals will degrade and recover, the method offers a way to refine the calibration strategies used by the experiment. This is not just a theoretical exercise; it has practical value for the thousands of scientists who rely on the Compact Muon Solenoid to make precise measurements of the universe. If the detector's sensors are clouded and uncorrected, the data they collect becomes unreliable. By using these predictive tools, the collaboration can ensure that the measurements remain sharp and accurate, even as the machine continues to operate in one of the most radiation-filled environments on Earth. The researchers plan to expand this work to include more crystals and explore even more advanced learning strategies, aiming to build a system that can handle the entire detector at once. For now, the success of this approach provides a solid foundation for keeping the world's most powerful particle camera in perfect focus.