Exploring phase transitions and dynamical processes in tropical moist convection using machine learning

Atmospheric precipitation in the tropics has previously been viewed as an example of self-organized criticality, with a critical value of the water vapor marking a continuous phase transition to a regime of strong convection and precipitation. Several recent studies have investigated phase transitions in statistical physics systems using data-driven dimensionality reduction techniques, and demonstrated that machine learning is able to recognize order and symmetry breaking, critical points, transition temperature ranges, and learn to differentiate first and second order phase transitions in an unsupervised manner. Here we investigate how these methods can be applied to tropical moist convection using simulations from a global storm resolving model that resolves sub-grid-scale deep convection, as part of the DYAMOND experiment.

Learned “latent” representations of high-resolution moisture fields show clear separation between precipitating and non-precipitating regimes. By investigating the spatial and temporal evolution of these latent space representations we investigate whether convection is initiated by thermodynamic fluctuations or larger scale dynamic processes. The time evolution of these latent representations indicates that latent-space transitions corresponding to greater Euclidean distances are more strongly correlated with transitions to regions of strong convection, supporting the hypothesis that the majority of convection is initiated by dynamic, rather than thermodynamic, processes. Using these methods, we investigate whether it is possible to derive a bottom-up order parameter that characterizes tropical moist convection as a continuous phase transition.