DataSheet1_Data-driven modeling of beam loss in the LHC.PDF

In the Large Hadron Collider, the beam losses are continuously measured for machine protection. By design, most of the particle losses occur in the collimation system, where the particles with high oscilla-tion amplitudes or large momentum error are scraped from the beams. The particle loss level is typically optimized manually by changing control parameters, among which are currents in the focusing and defocusing magnets. It is generally challenging to model and predict losses based only on the control parameters, due to the presence of various (non-linear) effects in the system, such as electron clouds, resonance effects, etc., and multiple sources of uncertainty. At the same time understanding the influence of control parameters on the losses is extremely important in order to improve the operation and performance, and future design of accelerators. Prior work [1] showed that modeling the losses as an instantaneous function of the control parameters does not generalize well to data from a different year, which is an indication that the leveraged statistical associations are not capturing the actual mechanisms which should be invariant from 1 year to the next. Given that this is most likely due to lagged effects, we propose to model the losses as a function of not only instantaneous but also previously observed control parameters as well as previous loss values. Using a standard reparameterization, we reformulate the model as a Kalman Filter (KF) which allows for a flexible and efficient estimation procedure. We consider two main variants: one with a scalar loss output, and a second one with a 4D output with loss, horizontal and vertical emittances, and aggregated heatload as components. The two models once learned can be run for a number of steps in the future, and the second model can forecast the evolution of quantities that are relevant to predicting the loss itself. Our results show that the proposed models trained on the beam loss data from 2017 are able to predict the losses on a time horizon of several minutes for the data of 2018 as well and successfully identify both local and global trends in the losses.

在大型强子对撞机（Large Hadron Collider）中，为保障机器安全，研究人员会持续监测束流损失情况。根据设计，绝大多数粒子损失都发生在准直系统中：该系统会刮除束流中振荡幅度过高或动量误差过大的粒子。束流损失水平通常通过调整控制参数手动优化，其中包括聚焦与散焦磁铁的励磁电流。仅依靠控制参数对损失进行建模与预测通常极具挑战，这是因为加速器系统中存在多种（非线性）效应，例如电子云、共振效应等，同时还存在多类不确定性来源。与此同时，明确控制参数对束流损失的影响，对于提升加速器的运行性能与未来的设计优化至关重要。已有研究[1]表明，将束流损失建模为控制参数的瞬时函数时，其泛化性无法适配不同年份的数据集，这意味着所利用的统计关联并未捕捉到跨年份应保持不变的真实物理机制。鉴于该问题大概率源于滞后效应，本文提出将束流损失建模为不仅包含瞬时控制参数，还涵盖过往观测到的控制参数与历史损失值的函数。通过标准的重参数化方法，我们将该模型重构为卡尔曼滤波（Kalman Filter, KF），从而实现灵活且高效的参数估计流程。本文考虑两种主要的模型变体：一种为单输出标量损失模型，另一种为四维输出模型，其输出分量包括束流损失、水平与垂直发射度，以及总热负载。完成训练后的两种模型均可对未来多个时间步长进行推演，而第二种模型还可预测与束流损失预测相关的物理量的演化趋势。实验结果表明，基于2017年束流损失数据训练得到的所提模型，可对2018年的数据集实现数分钟时间跨度的损失预测，并能准确识别束流损失的局部与全局趋势。