ABSTRACT

This paper examines the potential of reconstructing the Atlantic Meridional Overturning Circulation (AMOC) using surface data and linear machine learning algorithms. The algorithms are trained on pre-industrial control simulations with the aim of finding an algorithm that can reconstruct the AMOC robustly across multiple climate models. Predictors include a combination of surface temperature and surface salinity, as well as a combination of simultaneous and lagged values relative to the AMOC. For most climate models, the correlation skill of the AMOC reconstructions is greater than 0.7. This reconstruction model involves thousands of predictors and is therefore difficult to interpret. To improve interpretability, machine learning algorithms were applied to Laplacian eigenvectors, which are an orthogonal set of spatial patterns that can be ordered from largest to smallest spatial scale. The skill of the new algorithms is comparable to that based on gridded data, but the new algorithms have the advantage that dimension reduction can be more meaningfully interpreted. The most important predictors were simultaneous and lagged time series of area-averaged surface temperature, and a pattern that measures the east–west salinity difference over the basin surface lagged in time. These three predictors could recover a substantial fraction of the total skill from machine learning algorithms for most climate models.

摘要

本文研究了使用表面数据和线性机器学习算法重建大西洋经向翻转环流 (AMOC) 的潜力。这些算法在工业前控制模拟上进行了训练，目的是找到一种可以在多个气候模型中稳健地重建 AMOC 的算法。预测变量包括地表温度和地表盐度的组合，以及相对于 AMOC 的同时值和滞后值的组合。对于大多数气候模型，AMOC 重建的相关性大于 0.7。该重建模型涉及数千个预测变量，因此难以解释。为了提高可解释性，机器学习算法被应用于拉普拉斯特征向量，它们是一组正交的空间模式，可以从最大到最小的空间尺度进行排序。新算法的技巧与基于网格数据的算法相当，但新算法的优势在于可以更有意义地解释降维。最重要的预测因子是面积平均地表温度的同时和滞后时间序列，以及测量盆地表面东西盐度差异的模式滞后时间。对于大多数气候模型，这三个预测器可以从机器学习算法中恢复大部分技能。最重要的预测因子是面积平均地表温度的同时和滞后时间序列，以及测量盆地表面东西盐度差异的模式滞后时间。对于大多数气候模型，这三个预测器可以从机器学习算法中恢复大部分技能。最重要的预测因子是面积平均地表温度的同时和滞后时间序列，以及测量盆地表面东西盐度差异的模式滞后时间。对于大多数气候模型，这三个预测器可以从机器学习算法中恢复大部分技能。