High-Performance Modelling of a High-Intensity Cyclotron

本提案的目的是建立一个TRIUMF 500 MeV回旋加速器的在线模型。这可以通过开发一种全新的算法来实现，该算法将弥合现有快速包络码与用于模拟高强度回旋加速器的相对较慢的多粒子码之间的差距。为了进一步提高速度，我们还打算使用我们新开发的代码来训练一个神经网络，作为加速器的代理模型。在线模型正在改变我们在TRIUMF操作粒子加速器和光束线的方式。在线建模的关键是速度。模型需要在非常短的时间内运行，在一秒或更短的时间内，以允许模型、机器和人类操作员之间的实时交互。速度对于自动机器优化也是必不可少的，在收敛到最佳调谐之前可能会对模型进行大量调用。TRIUMF在线建模的主力是一个包络码，它跟踪粒子分布的第二时刻的演变，而不跟踪单个粒子。基于Sacherer开发的算法，该代码准确地考虑了光束内粒子之间的电磁相互作用。这对于精确模拟高强度机器是必不可少的。除其他外，该代码通常用于实时模拟TRIUMF高强度30 MeV电子直线加速器，并取得了巨大成功。TRIUMF的另一个高强度驱动加速器是500兆电子伏特回旋加速器。不幸的是，对于这样的机器，包络码是一个糟糕的模型。原因是根本的。另一种选择是使用多粒子代码，但是这种代码对于在线建模来说太慢了。一次对TRIUMF回旋加速器的模拟需要大约106个宏观粒子，并且需要大约10个小时才能完成（在单个处理器核心上）。为了将计算速度提高几个数量级，我们建议开发一种混合算法，其中在某些维度上使用统计矩，而在其他维度上使用类似于细胞内粒子的方法。我们小组的一名研究生最近推导出了一个类似的混合模型。我们打算将这项工作扩展到我们的回旋加速器上。如果我们在垂直方向上使用力矩，在其他两个方向上使用细胞内的粒子，我们将所需的宏观粒子数量减少了2/3，将106变为104，并可能将长达数小时的模拟变为几分钟的模拟。有了这个快速代码，就有可能运行足够多的模拟来训练神经网络，作为一个极快运行的回旋加速器代理模型。

The objective of this proposal is to develop an on-line model of TRIUMF's 500 MeV cyclotron. This can be achieved by developing a radically novel algorithm, that would bridge the gap between the existing fast envelope codes, and the relatively slow multiparticle codes used to model high-intensity cyclotrons. For an additional speed up, we also intend to use our newly developed code to train a neural network to serve as a surrogate model of the accelerators. On-line models are changing the way we operate particle accelerators and beamlines at TRIUMF. The key to on-line modelling is speed. The model needs to run in a very short time, on the order of a second or less, to allow real time interaction between the model, the machine, and the human operator. Speed is also essential for automatic machine optimization, which may make a large number of calls to the model before converging to an optimum tune. The workhorse of on-line modelling at TRIUMF is an envelope code, which tracks the evolution of the second moments of the particle distribution, without tracking individual particles. Based on the algorithm developed by Sacherer, this code accurately takes into account the electromagnetic interactions between particles within the beam. This is essential to accurately model high-intensity machines. Among other things, this code is used routinely to model, in real-time, the TRIUMF high-intensity 30 MeV electron linear accelerator, with great success. The other high-intensity driver accelerator at TRIUMF is the 500 MeV cyclotron. Unfortunately, an envelope code is a poor model for such a machine. The reasons are fundamental. The alternative is to use a multiparticle code, but such a code is orders of magnitude too slow for on-line modelling. One such simulation of the TRIUMF cyclotron requires in the order of 106 macro-particles, and takes on the order of 10 hours to complete (on a single processor core). To speed up the calculations by orders of magnitude, we propose to develop a hybrid algorithm, where statistical moments are used in some dimensions, while a particle-in-cell-like method is use in the others. A similar hybrid model has recently been derived by a graduate student working in our group. We propose to expand this work for application to our cyclotron. If we use moments in the vertical direction, and a particle-in-cell in the other two directions, we reduce the required number of macro-particles by a power 2/3, changing 106  into 104, and potentially turning hours-long simulations into minutes-long ones. With this fast code in hand, it will become possible to run a large enough number of simulations to train a neural network to serve as an extremely fast-running surrogate model of the cyclotron.