Figure 10.

Schematic of our hybrid convolution deep network. Top: our hybrid block. Bottom: An example model layout consisting of four blocks. The Input block to the hybrid networks are the profiles of winds and temperature. The Aux input refers to the surface pressure and geopotential, which are injected to each vertical layer of each block. Arrows indicate layers, e.g., a dense (fully connected) layer taking the inputs and aux inputs. Stack indicates the repeating of an object to enable concatenation with data that has vertical structure (e.g., wind profiles.).