权益分类	功能权益	普通用户	{{item.name}}会员
{{category.name}}	{{benefitItem.name}}

Machine Learning for Space Physics

空间物理机器学习

基本信息

批准号：
ST/T002255/1
负责人：
Sebastian Hoenig
金额：
$ 11.45万
依托单位：
University of Southampton
依托单位国家：
英国
项目类别：
Research Grant
财政年份：
2019
资助国家：
英国
起止时间：
2019 至无数据
项目状态：
已结题

来源：
https://gtr.ukri.org/projects?ref=ST%2FT002255%2F1
关键词：
Machine Learning Space Physics

项目摘要

Machine learning is a very hot topic in computer science these days. As a world we are generating ever greater volumes ofdata, and we need to find effective ways to gather and analyse that data, often by searching for regular patterns in datasets. The human eye is very good at picking out patterns either from images or from simple time series graphs. However,the human eye comes with its own biases: if you are trying to pick out blips in a single line trace on a screen your selectionmay not always be the same, but may depend on what has come before. Reproducibility is a huge issue here and onewhich impacts any kind of data science: if we are to do an experiment, or pick out interesting features from data, we want tomake sure we get the same result every time given the same initial input. Furthermore, as our input data streams getbigger and bigger, it is extremely time consuming (and a bit boring!) to look through all the data by eye to pick out the kindof features that we want. This is where the extremely powerful tool known as machine learning can help. In this work wepropose to use machine learning to pick out particular signatures from large catagloues of Space Physics data - but thecomputer analysis methods that we will develop will be applicable across multiple disciplines.The Space Physics problem we are interested in is called magnetic reconnection: it is a very energetic process which cantake place when two oppositely directed magnetic field lines meet, come together, and break. Right before reconnectionhappens the field lines are holding lots of energy, but as soon as they break this energy can be released into multiple formsincluding kinetic energy and thermal energy (heating). The field lines change shape after they break and these newly shapedfield lines can "ping" away from the site of reconnection, much like an elastic band that has been snapped. Thefield lines also carry with them charged particles, and these particles can heat up or change their flow direction as a resultof the transfer of energy.In the solar system everything happens on a giant scale, and magnetic reconnection can involve the magnetic field linesand plasma of the Sun and of several magnetised planets, including, but not limited to Mercury, Earth, Jupiter and Saturn.Spacecraft flying through the solar system have instruments which can measure magnetic fields and plasmas, and thuscan sample any changes associated with reconnection.The changes in the shape and orientation of magnetic fields and in the temperature and flow characteristics of chargedparticles can be observed by spacecraft. When scientists examine spacecraft data to search for evidence of thisreconnection process, they know what they are looking for in the field and plasma data. There is a huge amount ofspacecraft data: years and years' worth, with measurements taken several times a second. Reconnection can happenevery few minutes at some planets. It would be impossible for a human being to look through all the data and pick outevery time reconnection happened in our enormous catalogue.The purpose of this research is to teach the computer what reconnection signatures look like to a human eye, and to trainthe computer to pick these signatures out itself. This technique is called machine learning, and it has many advantages,because computers can be taught to work more quickly than humans, to give the same answer every time, and to not showbiases.The ultimate goal at the end of this project is to have trained the computer to select reconnection signatures, and to be ableto roll out this technique on multiple data sets from the solar system. This will be particularly useful for scientists who wantto conduct large studies of the behaviour of magnetic fields and plasma across the solar system, under different conditionsand over multiple years.

机器学习是当今计算机科学中的一个非常热门的话题。作为一个世界，我们正在产生越来越多的数据，我们需要找到有效的方法来收集和分析这些数据，通常是通过在数据集中搜索规则模式。人类的眼睛非常善于从图像或简单的时间序列图中挑选出模式。然而，人类的眼睛也有自己的偏见：如果你试图在屏幕上的单线轨迹中挑选光点，你的选择可能并不总是相同的，而是可能取决于之前出现的情况。再现性在这里是一个巨大的问题，它会影响到任何类型的数据科学：如果我们要做一个实验，或者从数据中挑选出有趣的特征，我们希望确保每次给定相同的初始输入，我们都会得到相同的结果。此外，随着我们的输入数据流变得越来越大，通过肉眼查看所有数据来挑选出我们想要的特征是非常耗时的（而且有点无聊！）。这就是被称为机器学习的强大工具可以提供帮助的地方。在这项工作中，我们建议使用机器学习从大型空间物理数据目录中挑选出特定的特征-但是我们将开发的计算机分析方法将适用于多个学科。我们感兴趣的空间物理问题被称为磁重联：这是一个非常有能量的过程，当两个方向相反的磁场线相遇，走到一起，然后断裂时就会发生。就在重新连接发生之前，磁场线持有大量能量，但一旦它们断开，这些能量可以以多种形式释放，包括动能和热能（加热）。磁场线在断裂后会改变形状，这些新形成的磁场线会从重新连接的位置“砰”地一声离开，就像一根松紧带被折断了一样。电场线也携带带电粒子，这些粒子可以加热或改变它们的流动方向作为能量转移的结果。在太阳系中，一切都发生在一个巨大的规模上，磁重联可能涉及太阳和几个磁化行星的磁力线和等离子体，包括但不限于水星、地球、木星和土星。穿越太阳系的航天器上有仪器可以测量磁场和等离子体，因此可以对与重联有关的任何变化进行采样。航天器可以观测到磁场形状和方向的变化，以及带电粒子的温度和流动特性的变化。当科学家们检查宇宙飞船的数据来寻找这种重新连接过程的证据时，他们知道他们在寻找磁场和等离子体数据。有大量的航天器数据：年复一年的价值，每秒进行几次测量。在一些行星上，重联每隔几分钟就会发生一次。人类不可能浏览所有的数据，并在我们庞大的目录中挑选出每次重新连接发生的时间。这项研究的目的是教会计算机人眼看到的重连签名是什么样子，并训练计算机自己挑选这些签名。这种技术被称为机器学习，它有很多优点，因为计算机可以被教导比人类更快地工作，每次都给出相同的答案，而且不会表现出偏见。这个项目的最终目标是训练计算机选择重新连接签名，并能够在来自太阳系的多个数据集上推出这项技术。这对于那些想要在不同条件和多年时间里对整个太阳系的磁场和等离子体的行为进行大规模研究的科学家来说尤其有用。