Automatic detection of ICMEs at 1 AU : a deep learning approach

Interplanetary Coronal Mass Ejections (ICMEs) are the interplanetarymanifestation of coronal mass ejections.Decades of studies through in situ measurements shed light on theirtypical characteristics : enhanced and smoothly rotating magneticfield, low proton temperature, declining velocity profile and lowplasma beta, etc. However, these features are not all observed for eachICME. In addition, they have a strong variability due to theirintrinsic nature, the way the spacecraft crosses the structure and thedifferent biases introduced by the observer. As a result, there is noreal consensus on how to identify an ICME, leading to disagreements inexisting catalogs which are poorly reproducible or extensible. In this work, we describe an automatic identification method based onconvolutional neural network trained on in situ measurements by theWIND spacecraft over the period 1997-2015. In addition to providing theobserver a fast and reproducible way to identify ICMEs, the algorithmfound about 300 new events. Working without prior knowledge about what ICMEs are other thanlabeled raw data, the method is quite robust and can be used in thefuture to identify signatures of other plasma structures and withmultiple spacecraft.