Using machine learning to unearth the aspects of deep convection that are robustly predictable from the local environmental state

NASA's upcoming INCUS mission will observe tropical deep convection at unprecedented spatiotemporal resolution (∼3 km horizontal sampling at intervals of 30 s, 90 s, and 120 s). INCUS will observe a wide variety of deep convective environments and tropical storms at different stages of their life cycle, producing many distinct convective outcomes. How do we distinguish these outcomes? Which differences are associated with genuinely different local environmental states, and which aspects vary even when the local environment is effectively the same?

We use Kernel Flows, a nonlinear machine learning estimator, to separate aspects of local deep convection that are constrained by their environmental state from those that are not. We first define the environmental state as a vector X formed from variables always present in the troposphere describing background thermodynamic and kinematic conditions, independent of whether convection is active (temperature, pressure, water vapor, horizontal wind fields). We define convective variables as vertical velocity, condensed water mass, and convective mass flux, which arise only where convection is present and correspond to quantities INCUS algorithms are designed to retrieve.

The analysis uses a database of high-resolution simulations (100 m horizontal grid spacing) across tropical and subtropical regions in anticipation of INCUS. From these simulations, we extract 25 km × 25 km neighborhoods containing a mixture of deep moist convective updrafts and surrounding non-updrafts. Within each neighborhood, we coarsen the environmental and convective variables to approximately 3 km resolution to match anticipated INCUS radar resolution. We represent the environmental state using principal components and define 19 scalar descriptors (Y_i) to characterize various aspects of convection.

Using Kernel Flows with no assumptions on Gaussian conditional distributions, we learn the functional relationship between each convective descriptor (Y_i) and environmental state (X) independently. We quantify how much variability in each descriptor is associated with environmental state differences by comparing the estimator's error variance to the prior variance of Y_i.

We find that aggregate convective descriptors exhibit variability almost entirely associated with environmental state differences. Total vertical velocity and total convective mass flux over the neighborhood, along with neighborhood-mean column maxima of these quantities, achieve R² ≥ 0.97 with at least 84% of standard deviation explained by the environmental state. These convective descriptors act as invariants of the local environment: differences in these metrics reflect environmental state (X) differences, rather than natural variability within an equivalent environmental state. Conversely, convective descriptors emphasizing horizontal and vertical organization of updrafts, such as how total updraft area is divided among horizontally contiguous clusters and maximum heights of vertical velocity and convective mass flux, have R² values typically below 0.5–0.55 and retain substantial variability even when X does not vary significantly. For these descriptors, only a modest fraction of variance is associated with environmental state differences, indicating that most observed differences reflect variability not captured by the local environmental state, potentially from smaller-scale dynamics or stochastic processes.

Our findings identify which aspects of deep convection are almost entirely tied to their local environmental state at spatiotemporal scales commensurate with INCUS observations.