Observed nonlinearities in the seasonal evolution of monsoons have been previously explained using theories for Hadley circulations driven by zonally symmetric thermal forcings, even though monsoonal forcings deviate strongly from the assumption of zonal symmetry. Here, an idealized model of a dry, three-dimensional atmosphere is used to compare the response to zonally symmetric and asymmetric off-equatorial thermal forcings. For symmetric forcings, the zonal-mean, cross-equatorial mass flux increases more rapidly with the amplitude of the forcing once the forcing becomes strong enough to reduce the upper-tropospheric absolute vorticity to near zero, consistent with previous studies of the transition to angular momentum–conserving flow. For zonally asymmetric forcings, the zonal-mean cross-equatorial flow exhibits a similar dependence on forcing strength and a similar reduction of the zonal-mean upper-level vorticity, but asymmetric forcings also produce strong zonal overturnings with subsidence west of the heating, as in the well-known linear response to off-equatorial heatings. The mass flux in these zonal overturnings increases linearly with forcing strength until its rate of increase tapers off for the strongest forcings; the total upward mass flux (i.e., the zonal-mean plus zonally asymmetric components) increases linearly with the strength of zonally asymmetric forcings and exhibits no abrupt or nonlinear dependence on forcing amplitude. These results indicate the importance of considering the zonally asymmetric part of the divergent response to off-equatorial forcings and suggest that theories based on zonally symmetric forcings need further examination before they can be assumed to describe observed monsoons.

Shallow meridional overturning circulations are superimposed on the deep circulations that produce precipitation in nearly all monsoon regions, and these shallow circulations transport subtropical, mid‐tropospheric dry air into the tropical monsoon precipitation maxima. Here horizontal moisture advection produced by shallow meridional circulations is characterized in the monsoon regions of West Africa, South Asia, Australia, and southern Africa during local summer. Horizontal flow in the upper and lower branches of the shallow meridional circulations consistently dries and moistens air, respectively, in the continental precipitation maxima of each region. The peak drying by horizontal advection occurs at a lower altitude than peak winds in the upper branch of the shallow circulations, consistent with the small scale height of water vapour. Advection of time‐mean moisture by time‐mean wind dominates horizontal moisture advection in South Asia and West Africa, while most horizontal moisture advection in Australia and southern Africa is produced by transient eddies. Much of the transient eddy advection can be accurately represented as a first‐order horizontal diffusion with a constant, globally uniform diffusivity. These results suggest that horizontal moisture advection in theoretical and conceptual models of seasonal mean monsoons can be adequately represented in terms of time‐mean winds plus a simple horizontal moisture diffusion. Finally, interannual variations in the summer mean regional averages of monsoon precipitation and horizontal advective drying in the lower free troposphere are shown to be negatively correlated in most regions, consistent with the hypothesis that advective drying by shallow meridional circulations inhibits monsoon precipitation.

Arctic sea ice motions can be modeled using a CNN with predictors of previousday ice velocity, concentration and present-day surface wind.• The superiority of CNN over baseline models suggests the importance of non-local connections compared to local point-wise interactions.• The success of the CNN model of ice motion suggests potential for combining machine learning with physics-based models to simulate sea ice.

Arctic sea ice concentration is often coarsely observed and numerically computed despite its importance for polar climate system. In this work we present three machinelearning methods to recover the original high-resolution images from the coarse-grained low-resolution counterparts. The promising results indicate a possibility of extending the application to a broad range of geophysical variables.