Next Generation, Physics-Inspired AI for Space Weather Forecasting

空间天气描述了近地空间条件的变化。太空天气影响社会的主要方式之一是通过在地面电网和管道中产生异常电流（称为地磁感应电流或gic）。这些gic可以加速系统的老化，或者更严重的是导致电力变压器等组件立即失效。这项研究将在理解和预测我们在地面上何时有遭受大规模gic风险方面取得飞跃。全球变暖是由地球磁场的快速变化驱动的，在近地空间有一系列的现象，但其中最重要的是磁层亚暴。在亚暴期间，地球磁场和入射太阳风之间的相互作用导致能量的转移。这些额外的能量主要储存在等离子体和磁场能量中，在行星的夜侧，一个被称为磁尾的区域。能量被储存，直到系统达到稳定的极限，在这一点上，能量被爆炸性地释放，再次通过磁重联的过程。这导致了可观测的现象，如极光。然而，这一过程可能会产生可怕的太空天气后果，造成极端的电离层电流，并对卫星和其他基础设施构成风险，然而，即使是我们最先进的方法也难以预测它何时会发生。当对磁层亚暴开始时的无数零星和局部过程知之甚少时，理解和预测磁场变化是一个非常困难的问题。其中一个基本问题是该系统的规模。所涉及的过程是零星的和局部的，它们可以操作的领域是巨大的。这项奖学金的目的是了解磁层变得不稳定的过程和不稳定性，并利用它来产生尖端的，物理启发的空间天气预报模型。我将使用“大数据”技术准确而稳健地处理来自地球上几个任务的大量数据，以表征和预测亚暴可能发生的条件。我将开发贝叶斯蒙特卡罗方法来估计它们的空间和时间尺度并确定因果关系。然后，我将利用这种理解来生成先进的机器学习模型，预测亚暴发生的时间和地点，以及极光椭圆形的属性和位置。然后，我将把这些组合在一起，创建一个受物理学启发的预测地磁扰动的模型。这对于提供准确和可靠的预测区域何时面临危险的gic风险是必要的。受物理启发的过程将确保对极端条件的模型外推比“黑匣子”外推更可靠。在这个奖学金的过程中，我将与世界领先的等离子体稳定性（MSSL）和磁尾动力学（密歇根）的专家合作，利用尖端的全球模型（密歇根）来告知最先进的机器学习模型。然后，我将为利益相关者创建稳健可靠的模型（英国气象局）。

Space weather describes the variability of conditions in near-Earth space. One of the primary ways in which space weather can impact society is through the generation of anomalous currents (termed Geomagnetically Induced Currents, or GICs) in power networks and pipelines on the ground. These GICs can accelerate the ageing of systems, or more critically lead to the immediate failure of components such as power transformers. This research will take a leap forward in understanding and predicting when we are at risk of suffering large GICs on the ground.GICs are driven by rapid changes in the Earth's magnetic field, and there are a range of phenomena in near-Earth space that are responsible, but one of the most important is the magnetospheric substorm. During a substorm, interactions between the magnetic field of the Earth and the incident solar wind results in the transfer of energy. This additional energy is principally stored in plasma and magnetic field energy on the nightside of a planet in a region known as the magnetotail. Energy is stored until the system reaches the limit of stability, at which point the energy is explosively released, again through the process of magnetic reconnection. This leads to observable phenomena such as the aurora. However, this process can have dire space weather consequences, causing extreme ionospheric currents and posing risks to satellites and other infrastructure, yet even our most sophisticated methods struggle to predict when it will occur.Understanding and forecasting magnetic field variability is a hugely difficult problem when the myriad of sporadic and localised processes at the start of a magnetosphere substorm are poorly understood. One of the fundamental issues is the scale of the system. The processes involved are sporadic and localised, and the domain in which they could operate is huge. The aim of this fellowship is to understand the processes and instabilities by which the magnetosphere becomes unstable, and use this to generate cutting-edge, physics-inspired space weather forecasting models.I will accurately and robustly process huge volumes of data from several missions at the Earth using 'big data' techniques to characterize and predict the conditions under which the substorm is likely to occur. I will develop Bayesian Monte Carlo methods to estimate their spatial and temporal scales and determine causality. I will then use this understanding to generate cutting-edge machine learning models of when and where substorms will occur, as well as the properties and location of the auroral oval. I will then put this together to create a physics-inspired model of forecasting geomagnetic perturbations. This is necessary to provide precise and reliable predictions of when regions are at risk of dangerous GICs. The physics-inspired process will ensure that the model extrapolations to extreme conditions are more reliable than 'black box' extrapolations.During the course of this fellowship I will collaborate with world leading experts on plasma stability (MSSL) and magnetotail dynamics (Michigan), utilizing cutting edge global models (Michigan) to inform state-of-the-art machine learning models. I will then create robust and reliable models for the benefit of stakeholders (Met Office).