Representing unresolved moist convection in coarse‐scale climate models remains one of the main bottlenecks of current climate simulations. Many of the biases present with parameterized convection are strongly reduced when convection is explicitly resolved (i.e., in cloud resolving models at high spatial resolution approximately a kilometer or so). We here present a novel approach to convective parameterization based on machine learning, using an aquaplanet with prescribed sea surface temperatures as a proof of concept. A deep neural network is trained with a superparameterized version of a climate model in which convection is resolved by thousands of embedded 2‐D cloud resolving models. The machine learning representation of convection, which we call the Cloud Brain (CBRAIN), can skillfully predict many of the convective heating, moistening, and radiative features of superparameterization that are most important to climate simulation, although an unintended side effect is to reduce some of the superparameterization's inherent variance. Since as few as three months' high‐frequency global training data prove sufficient to provide this skill, the approach presented here opens up a new possibility for a future class of convection parameterizations in climate models that are built “top‐down,” that is, by learning salient features of convection from unusually explicit simulations.

Many theories for the Madden‐Julian oscillation (MJO) focus on diabatic processes, particularly the evolution of vertical heating and moistening. Poor MJO performance in weather and climate models is often blamed on biases in these processes and their interactions with the large‐scale circulation. We introduce one of the three components of a model evaluation project, which aims to connect MJO fidelity in models to their representations of several physical processes, focusing on diabatic heating and moistening. This component consists of 20 day hindcasts, initialized daily during two MJO events in winter 2009–2010. The 13 models exhibit a range of skill: several have accurate forecasts to 20 days lead, while others perform similarly to statistical models (8–11 days). Models that maintain the observed MJO amplitude accurately predict propagation, but not vice versa. We find no link between hindcast fidelity and the precipitation‐moisture relationship, in contrast to other recent studies. There is also no relationship between models' performance and the evolution of their diabatic heating profiles with rain rate. A more robust association emerges between models' fidelity and net moistening: the highest‐skill models show a clear transition from low‐level moistening for light rainfall to midlevel moistening at moderate rainfall and upper level moistening for heavy rainfall. The midlevel moistening, arising from both dynamics and physics, may be most important. Accurately representing many processes may be necessary but not sufficient for capturing the MJO, which suggests that models fail to predict the MJO for a broad range of reasons and limits the possibility of finding a panacea.

Implementing artificial neural networks is commonly achieved via high-level programming languages such as Python and easy-to-use deep learning libraries such as Keras. These software libraries come preloaded with a variety of network architectures, provide autodifferentiation, and support GPUs for fast and efficient computation. As a result, a deep learning practitioner will favor training a neural network model in Python, where these tools are readily available. However, many large-scale scientific computation projects are written in Fortran, making it difficult to integrate with modern deep learning methods. To alleviate this problem, we introduce a software library, the Fortran-Keras Bridge (FKB). This two-way bridge connects environments where deep learning resources are plentiful with those where they are scarce. The paper describes several unique features offered by FKB, such as customizable layers, loss functions, and network ensembles. The paper concludes with a case study that applies FKB to address open questions about the robustness of an experimental approach to global climate simulation, in which subgrid physics are outsourced to deep neural network emulators. In this context, FKB enables a hyperparameter search of one hundred plus candidate models of subgrid cloud and radiation physics, initially implemented in Keras, to be transferred and used in Fortran. Such a process allows the model’s emergent behavior to be assessed, i.e., when fit imperfections are coupled to explicit planetary-scale fluid dynamics. The results reveal a previously unrecognized strong relationship between offline validation error and online performance, in which the choice of the optimizer proves unexpectedly critical. This in turn reveals many new neural network architectures that produce considerable improvements in climate model stability including some with reduced error, for an especially challenging training dataset.