Search

We present a machine learning based emulator of a microphysics scheme for condensation and precipitation processes (Zhao‐Carr) used operationally in a global atmospheric forecast model (FV3GFS). Our tailored emulator architecture achieves high skill (≥94%) in predicting condensate and precipitation amounts and maintains low global‐average bias (≤4%) for 1 year of continuous simulation when replacing the Fortran scheme. The stability and success of this emulator stems from key design decisions. By separating the emulation of condensation and precipitation processes, we can better enforce physical priors such as mass conservation and locality of condensation, and the vertical dependence of precipitation falling downward, using specific network architectures. An activity classifier for condensation imitates the discrete‐continuous nature of the Fortran microphysics outputs (i.e., tendencies are identically zero where the scheme is inactive, and condensate is zero where clouds are fully evaporated). A temperature‐scaled conditional loss function ensures accurate condensate adjustments for a high dynamic range of cloud types (e.g., cold, low‐condensate cirrus clouds or warm, condensate‐rich clouds). Despite excellent overall performance, the emulator exhibits some deficiencies in the uppermost model levels, leading to biases in the stratosphere. The emulator also has short episodic skill dropouts in isolated grid columns and is computationally slower than the original Fortran scheme. Nonetheless, our challenges and strategies should be applicable to the emulation of other microphysical schemes. More broadly, our work demonstrates that with suitable physically motivated architectural choices, ML techniques can accurately emulate complex human‐designed parameterizations of fast physical processes central to weather and climate models. Plain Language Summary: In this study, we create computer code that uses machine learning to mimic a weather model's algorithm for handling how clouds form and rain falls. When used in the weather model to replace this algorithm, our machine learning code is highly accurate in simulations for a whole year. We achieve this by making smart code design choices. We split the code into two parts: one for cloud formation and one for rain and snow. This allows us to better build important aspects of these processes into the machine learning approach. For instance, clouds form where it is moist and evaporate when it gets dry, and rain and snow fall downward. Our code learns cloud behavior based on temperature to ensure it works both for cold, thin clouds high up in the sky and warm, thick clouds closer to the ground. Our work shows a path for suitably designed machine learning code to eventually replace important parts of weather and climate models, but also that this path still requires careful human design respecting known physical principles. Key Points: We build an emulator to replace the Zhao‐Carr Fortran microphysics scheme in FV3GFSThe integrated emulator sustains high skill throughout a 1‐year simulationTailoring the ML architecture to the structure of the underlying scheme greatly improves the online behavior of the emulator [ABSTRACT FROM AUTHOR]

Parameterization of subgrid‐scale processes is a major source of uncertainty in global atmospheric model simulations. Global storm‐resolving simulations use a finer grid (less than 5 km) to reduce this uncertainty by explicitly resolving deep convection and details of orography. This study uses machine learning to replace the physical parameterizations of heating and moistening rates, but not wind tendencies, in a coarse‐grid (200 km) global atmosphere model, using training data obtained by spatially coarse‐graining a 40‐day realistic geography global storm‐resolving simulation. The training targets are the three‐dimensional fields of effective heating and moistening rates, including the effect of grid‐scale motions that are resolved but imperfectly simulated by the coarse model. A neural network is trained to predict the time‐dependent heating and moistening rates in each grid column using the coarse‐grained temperature, specific humidity, surface turbulent heat fluxes, cosine of solar zenith angle, land‐sea mask and surface geopotential of that grid column as inputs. The coefficient of determination R2 for offline prediction ranges from 0.4 to 0.8 at most vertical levels and latitudes. Online, we achieve stable 35‐day simulations, with metrics of skill such as the time‐mean pattern of near‐surface temperature and precipitation comparable or slightly better than a baseline simulation with conventional physical parameterizations. However, the structure of tropical circulation and relative humidity in the upper troposphere are unrealistic. Overall, this study shows potential for the replacement of human‐designed parameterizations with data‐driven ones in a realistic setting. Plain Language Summary: Numerical models used for projecting climate change impacts must use ad‐hoc assumptions about the effects of unresolved small‐scale processes. These assumptions contribute to uncertainty in predicting how rainfall and temperature will change in the future. Expensive fine‐grid simulations which eliminate the need for some of these assumptions are possible to run for shorter (month‐to year‐long) duration. We use such a simulation to train a data‐driven representation of the effects of processes, like clouds, which are poorly simulated by a cheaper coarse‐grid model. The data‐driven representation (a neural network) predicts rates of temperature and moisture change in each column using inputs from that grid column. This approach has been previously shown to work for models with idealized boundary conditions, but not for the realistic setting we use. When this neural network is used in a coarse‐resolution model, the realism of many global skill metrics is as good or better than a baseline model with traditional representation of small‐scale processes. However, some features are degraded, such as the time‐evolving pattern of rainfall in the tropics and humidity in the upper atmosphere. This work is a first step toward the use of data‐driven representations of unresolved processes in realistic global atmospheric models. Key Points: Effective sources of heat and moisture are computed from a global storm‐resolving simulation accounting for semi‐resolved dynamicsA neural network is trained to predict columns of the effective sources using profiles of temperature and specific humidityWhen used online, stable month‐long simulations are possible although skill is not yet comparable to a previous corrective approach [ABSTRACT FROM AUTHOR]