We propose a framework leveraging machine learning to enable the development of Earth System Digital Twins (ESDTs) through a machine-learned (ML) stochastic bin microphysics (SBM) emulator. To achieve the fast observational feedback required by an ESDT, we plan to first implement a conditional variational auto-encoder/decoder (cVAE/D) model as a radar observation parameter (ROP) mapper that maps the high-dimensional hydrometeor size spectra produced by the physics-based SBM model to a manifold of much-reduced dimensionality, from which the original ROP mapper input can be accurately recovered efficiently when needed. The encoder of the cVAE/D projects the physics-based SBM outcomes, including the hydrometeor species and their size spectra, to a compressed manifold of much lower dimensionality than that of the original SBM model outcomes. Radar observation parameters (ROPs) for simulating radar signals will be part of the compressed manifold as prescribe variables. The corresponding decoder can invert the projection and reproduce the SBM model outcome. ROPs from the ROP mapper enable efficient radar observable calculations for feedback to the SBM emulator and, in turn, the host dynamic cloud-resolving model (CRM). More crucially, we will train a neural network (NN) as an SBM emulator, based on the Long Short-Term Memory (LSTM) model, to mimic the functioning of the physics-based SBM model through the manifold variables obtained from the ROP mapper for time-series prediction. We plan to use high temporal resolution (i.e., time steps of 1 or 2 minutes) physics-based SBM model outcomes within CRM simulations to train the LSTM SBM emulator. Our goal is to incorporate the SBM emulator into the CRM as a computationally efficient alternative to the expensive physics-based SBM model in Phase II of this proposed project if the feasibility of this approach is verified in Phase I. Anticipated
Benefits: NASA has considerable efforts involving weather and climate modeling for research and climate prediction, which would directly benefit. Researchers of weather products usually apply them in operational contexts, such as mission and field campaign planning and execution, which have demanding real-, near-real-, or in-time requirements. Our proposed ML-based microphysics emulation would support improved rapid responses in these regimes by allowing (real-time) observation feedback to be brought to bear that is currently infeasible or impractical. The global weather forecasting market was about $2.2 billion/year in 2021, projected to reach almost $4 billion/year by 2030. Our natural client base is the plethora of weather and climate analysis firms serving application areas, including planning, monitoring, and analysis supporting agriculture, fisheries, disaster planning/response, insurance, energy, and transportation, among others.
