Assessing the Potential of Deep Learning for Emulating Cloud Superparameterization in Climate Models With Real‐Geography Boundary Conditions

Abstract

Although global atmospheric model simulations are increasingly high-resolution, even under optimistic scenarios of enhanced computing performance, physically resolving the atmospheric turbulence controlling clouds will likely not be feasible for decades. Current climate model horizontal grid cells are typically 50-100 km wide but the turbulent updrafts governing cloud formation occur on scales of just tens to hundreds of meters and the microphysical processes regulating convection occur down at the micrometer Abstract We explore the potential of feed-forward deep neural networks (DNNs) for emulating cloud superparameterization in realistic geography, using offline fits to data from the superparameterized community atmospheric model. To identify the network architecture of greatest skill, we formally optimize hyperparameters using ∼250 trials. Our DNN explains over 70% of the temporal variance at the 15-min sampling scale throughout the mid-to-upper troposphere. Autocorrelation timescale analysis compared against DNN skill suggests the less good fit in the tropical, marine boundary layer is driven by neural network difficulty emulating fast, stochastic signals in convection. However, spectral analysis in the temporal domain indicates skillful emulation of signals on diurnal to synoptic scales. A closer look at the diurnal cycle reveals correct emulation of land-sea contrasts and vertical structure in the heating and moistening fields, but some distortion of precipitation. Sensitivity tests targeting precipitation skill reveal complementary effects of adding positive constraints versus hyperparameter tuning, motivating the use of both in the future. A first attempt to force an offline land model with DNN emulated atmospheric fields produces reassuring results further supporting neural network emulation viability in realgeography settings. Overall, the fit skill is competitive with recent attempts by sophisticated Residual and Convolutional Neural Network architectures trained on added information, including memory of past states. Our results confirm the parameterizability of superparameterized convection with continents through machine learning and we highlight the advantages of casting this problem locally in space and time for accurate emulation and hopefully quick implementation of hybrid climate models.Plain Language Summary Machine learning methods have been previously used to replace parameterizations (approximations) of atmospheric convection under very idealized scenarios (aqua-planets). The hope is that these machine learning emulators can help power the next generation of climate models with similar accuracy but at a fraction of the computational cost. But important questions remain about how learnable more realistic convection (over both land and ocean) is. Recently, the first attempt at machine learning replicated convection was made under these Earth-like conditions. But it required a highly specialized neural network as well as memory of the previous behavior of the atmosphere. This design would make ...

Non-Linear Dimensionality Reduction with a Variational Encoder Decoder to Understand Convective Processes in Climate Models

“…We selected the global sub-grid heating and moistening fields at 700 hPa for the evaluation of the VAE's R 2 (Figure 3). We choose dq/dt and dT/dt fields at this pressure level because of the limited skill in fitting lower tropospheric convective processes with neural nets that has been reported across multiple investigations, and which has been speculated to be associated with an underrepresentation of stochastic variability linked to shallow and deep convection (Gentine et al (2018); ; Mooers et al (2021); Wang et al (2021)). Both networks, VAE and reference ANN, exhibit similar emulation skill patterns for heating and moistening tendencies, including the skill deficits for low-level moistening tendencies in the tropics, as seen in previous studies.…”

“…Similar advances in the prognostic skill, where the ML approach is coupled to the dynamical core of a circulation model, of global precipitation distributions on an aquaplanet were achieved by Yuval and O'Gorman (2020), with random forests and neural networks. Beyond aquaplanets, Mooers et al (2021) showed that feed-forward neural nets also skilfully reproduce the superparameterization (SP) in the presence of real topography based on offline tests, where the ML approach is evaluated against test data, without implementing the resulting representation back into the GCM. Han et al (2020) likewise demonstrated the potential to learn SP with residual neural nets based on real topography data.…”

Correcting Coarse‐Grid Weather and Climate Models by Machine Learning From Global Storm‐Resolving Simulations

“…This makes the ML training problem easier to precisely formulate, but also sidesteps important real‐world complications such as orography, land surface heterogeneity, complex coastlines, etc. Other studies have tackled real‐world geography but only demonstrated offline or single‐column ML skill (Han et al., 2020; Mooers et al., 2021). Wang et al.…”

Non‐Linear Dimensionality Reduction With a Variational Encoder Decoder to Understand Convective Processes in Climate Models

“…Beyond aquaplanets, Mooers et al. (2021) showed that feed‐forward neural nets also skilfully reproduce the SuperParameterization (SP) in the presence of real topography based on offline tests, where the ML approach is evaluated against test data, without implementing the resulting representation back into the GCM. Han et al.…”

“…Similar advances in the prognostic skill, where the ML approach is coupled to the dynamical core of a circulation model, of global precipitation distributions on an aquaplanet were achieved by Yuval and O'Gorman (2020), with random forests and neural networks. Beyond aquaplanets, Mooers et al (2021) showed that feed-forward neural nets also skilfully reproduce the SuperParameterization (SP) in the presence of real topography based on offline tests, where the ML approach is evaluated against test data, without implementing the resulting representation back into the GCM. Han et al (2020) likewise demonstrated the potential to learn SP with residual neural nets based on real topography data.…”