However, MWNet has yet to be suitable for operational use by forecasters. Focusing on the quantitative performance represented by Mean Squared Error resulted in blurry predictions that entirely miss the critical high-frequency features of weather measurements. Furthermore, forecasters are naturally suspicious of predictions not backed by laws of physics, no matter their accuracy.
This work presents PhyDGAN, a physically constrained recurrent neural network trained in the GAN framework. PhyDGAN, trained for precipitation prediction for up to an hour into the future on a radar reflectivity dataset, disentangles base physical dynamics, represented by the advection-diffusion partial differential equation, from other residual factors that are unknown given the limited 2D radar reflectivity input data. Among other things, these residual factors are responsible for modeling high-frequency features.
PhyDGAN solves both aforementioned challenges, producing highly detailed and partially explainable probabilistic nowcasts. We quantify these improvements and present an analysis of PhyDGAN's probabilistic properties and performance trade-offs associated with the proposed changes.
Finally, we explore latent features of PhyDGAN and the learned disentanglement of physical and residual dynamics, shedding light on the reasoning behind the predictions and creating a novel advection field output solely from radar reflectivity data. PhyDGAN is operationally used by forecasters in Meteopress.

