Untangling Natural Aerosol Processes in Polar Regions by Implementing Novel Machine Learning Techniques

准确的气候模型对于提高我们对日益紧迫的人为气候变化问题的理解至关重要。气溶胶与云的相互作用仍然是气候建模中最大的不确定性来源之一（Myhre等人，2013年），限制了对气候变化的理解和改进气候预测的能力。为了改善气溶胶在气候模式中的表现，有必要了解天然气溶胶如何缓冲云对人为排放的敏感性。目前在原始地区（如极地）的自然气溶胶观测可用于限制GCMs（大气环流模式）和ESMs（地球系统模式）中工业化前气溶胶的参数化。然而，仍然缺乏对运输过程中控制气溶胶源和汇的物理过程的认识。拉格朗日方法的使用是向前迈出的一步，在拉格朗日方法中，使用轨迹来跟踪从气溶胶测量站返回的气团的路径。气象变量和潜在来源可与这些轨迹并列，然后结合气溶胶测量进行分析，以确定和了解天然气溶胶来源。由于需要考虑大量变量以及它们之间的相互作用，这一领域的进展受到阻碍；因此，对于传统的统计方法来说，情况太复杂了。为了能够解开和解释这些过程，该项目将探索新的机器学习技术，重点是卷积神经网络和长短期记忆架构递归神经网络。极地地区将是重点区域，不仅因为原始的环境，还因为这些地区的气候影响很大。北极对与复杂气候反馈相关的辐射收支扰动非常敏感（Pithan和Mauritsen， 2014）。南极半岛也是地球上变暖最快的地区之一，导致了当地冰冻圈的变化（Turner et al., 2005）。气候模式对该地区云层水平的预测仍然不足，这限制了我们对地表融化提供准确预测的能力，从而限制了气候系统的反馈。为了提高该地区云的模型表征，需要对天然气溶胶源有更深入的了解（King et al., 2015）。下面的气候模式预测了北极和南极的气溶胶水平，特别是较大的气候活性气溶胶。与观测结果相比，冬季期间气溶胶的缺失表明气候模式中缺少一个源（Frey等，2015）。两极的气溶胶主要由海盐颗粒组成。而海盐气溶胶浓度的峰值出现在冬季，与海冰面积最大的季节相吻合。最近有直接证据表明，海冰上方吹来的雪和升华是这些气溶胶的来源（Frey et al., 2019）。以前还没有将吹雪的参数化应用到GCM或ESM中，因此在当前模式中可以忽略关键的极地气候反馈。作为该项目的一部分开发的新框架，利用拉格朗日框架中的神经网络，将提高对两极天然气溶胶来源的相对影响的理解。该分析将确定吹雪是否是该地区重要的气溶胶来源，以及通过吹雪机制对气溶胶产生的主要影响。一旦这一点得到探索，吹雪的参数化（来自英国南极调查局）将首次被应用到ESM（英国气象局的UKESM）中，以调查气候系统中由此产生的反馈。这项研究将进一步了解自然气溶胶过程并改进气候预测。

Accurate climate models are essential to improve our understanding of the ever pressing issue of anthropogenic climate change. Aerosol-cloud interactions continue to be one of the largest sources of uncertainty in climate modelling (Myhre et al., 2013), limiting the understanding of climate change and ability to improve climate predictions. In order to improve aerosol representation in climate models it is essential to understand how natural aerosols might buffer the sensitivity of clouds to anthropogenic emissions. Present-day natural aerosol observations in pristine regions, like the poles, can be used to constrain pre-industrial aerosol parameterisations in GCMs (General Circulation Models) and ESMs (Earth System Models). However, there remains a lack of knowledge of the physical processes that control aerosol sources and sinks during transport.A step forward has been the use of a Lagrangian approach, in which trajectories are used to track the passage of air masses back from an aerosol measuring station. Meteorological variables and potential sources can be collocated to these trajectories and then analysed in conjunction with the aerosol measurements to identify and understand natural aerosol sources. Progress is hindered in this area due to the vast number of variables to consider, as well as interactions between them; the picture is therefore too complex for traditional statistical methods. In order to be able to untangle and interpret these processes this project will explore novel machine learning techniques, focusing on Convolutional Neural Networks and Long-Short Term Memory architecture recurrent neural networks. Polar Regions will be the area of focus, not only due to the pristine environment, but also because of the high climatic impact of these regions. The Arctic is very sensitive to perturbations in the radiative budget associated with complex climate feedbacks (Pithan and Mauritsen, 2014). The Antarctic peninsula is also one of the most rapidly warming areas on the planet, leading to changes in the local cryosphere (Turner et al., 2005). Climate models continue to under predict the level of cloud in this region, limiting our ability to provide accurate predictions for surface melt and therefore feedbacks in the climate system. In order to improve model representation of clouds in this region, there needs to be a greater understanding of natural aerosol sources (King et al., 2015).Climate models under predict aerosol levels in both the Arctic and Antarctic, particularly the larger climate-active aerosols. The lack of aerosols during the winter period in comparison to observations is indicative of a missing source in climate models (Frey et al., 2015). Aerosols at the poles mostly consist of sea salt particles. However, the peak in sea salt aerosol concentration is during the winter period, coinciding with the season of largest sea ice extent. There has recently been direct evidence to suggest that snow blowing above sea-ice and sublimating is a source of these aerosols (Frey et al., 2019). A parametrisation of blowing snow has not been implemented into a GCM or ESM before, so crucial polar climate feedbacks could be neglected in current models. The novel framework developed as part of this project, utilising neural networks in a Lagrangian framework, will improve understanding of the relative influence of sources of natural aerosols at the poles. This analysis will identify if blowing snow is a significant aerosol source in the region and the main influences on aerosol production through the blowing snow mechanism. Once this is explored, a parameterisation of blowing snow (from the British Antarctic Survey) will be implemented into an ESM (the Met Office's UKESM) for the first time to investigate the resultant feedbacks in the climate system. This research will further the understanding of natural aerosol processes and improve climate predictions.