Data driven non-equilibrium moist phase exchanges for atmospheric convection within a discontinuous Galerkin model of the compressible Euler equations

A neural network is trained to learn the mass exchanges between vapour, liquid and ice phases in atmospheric convection.

The network is trained on convection resolving output from a regional configuration of the LFRic model with a multi-moment microphysics parameterisation (CASIM).

The loss function for learning these phase exchanges is formulated under the assumptions of thermal and mechanical equilibrium (same temperature and pressure for all phases), and mechanical dis-equilibrium (different Gibbs free energies for all phases).

The network outputs determine the exchanges of vapour, liquid and ice so as to conserve mass, and the resulting change in entropy is determined from the network outputs so as to conserve energy.

The neural network is implemented in a thermodynamically consistent manner within a 2D vertical slice discontinuous Galerkin model of a moist, non-hydrostatic atmosphere in order to simulate the formation of three-phase clouds for convection at sub-km resolution.

The results are compared to those from a physics based representation of three-phase moist processes at thermodynamic equilibrium.