Global kilometer-scale models are the future of Earth system models as they can explicitly simulate organized convective storms and their associated extreme weather. Here, we comprehensively examined tropical mesoscale convective system (MCS) characteristics in the DYAMOND (DYnamics of the atmospheric general circulation modeled on non-hydrostatic domains) models for both summer and winter phases by applying eight different feature trackers to the simulations and satellite observations. Although different trackers produce substantial differences (a factor of 2-3) in observed MCS frequency and their contribution to total precipitation, model-observation differences in MCS statistics are more consistent among the trackers. DYAMOND models are generally skillful in simulating tropical mean MCS frequency, with multi-model mean biases of 2.9% over land and -0.5% over ocean. However, most models underestimate the MCS precipitation amount (23%) and their contribution to total precipitation (17%) relative to observations. These biases show large inter-model variability, but are generally smaller over land (13%) than over ocean (21%) on average. MCS diurnal cycle and cloud shield characteristics are better simulated than precipitation. Most models overestimate MCS precipitation intensity and underestimate stratiform rain contribution (up to a factor of 2), particularly over land. Models also predict a wide range of precipitable water in the tropics compared to reanalysis and satellite observations, and many models simulate a greater sensitivity of MCS precipitation intensity to precipitable water. The MCS metrics developed in this work provide process-oriented diagnostics for future model development efforts.

This is a test-case study assessing the ability of deep learning methods to generalize to a future climate (end of 21st century) when trained to classify thunderstorms in model output representative of the present-day climate. A convolutional neural network (CNN) was trained to classify strongly-rotating thunderstorms from a current climate created using the Weather Research and Forecasting (WRF) model at high-resolution, then evaluated against thunderstorms from a future climate, and found to perform with skill and comparatively in both climates. Despite training with labels derived from a threshold value of a severe thunderstorm diagnostic (updraft helicity), which was not used as an input attribute, the CNN learned physical characteristics of organized convection and environments that are not captured by the diagnostic heuristic. Physical features were not prescribed but rather learned from the data, such as the importance of dry air at mid-levels for intense thunderstorm development when low-level moisture is present (i.e., convective available potential energy). Explanation techniques also revealed that thunderstorms classified as strongly rotating are associated with learned rotation signatures. Results show that the creation of synthetic data with ground truth is a viable alternative to human-labeled data and that a CNN is able to generalize a target using learned features that would be difficult to encode due to spatial complexity. Most importantly, results from this study show that deep learning is capable of generalizing to future climate extremes and can exhibit out-of-sample robustness with hyperparameter tuning in certain applications.

Extreme precipitation events, including those associated with weather fronts, have wide-ranging impacts across the world. Here we use a deep learning algorithm to identify weather fronts in high resolution Community Earth System Model (CESM) simulations over the contiguous United States (CONUS), and evaluate the results using observational and reanalysis products. We further compare results between CESM simulations using present-day and future climate forcing, to study how these features might change with climate change. We find that detected front frequencies in CESM have seasonally varying spatial patterns and responses to climate change and are found to be associated with modeled changes in large scale circulation such as the jet stream. We also associate the detected fronts with precipitation and find that total and extreme frontal precipitation mostly decreases with climate change, with some seasonal and regional differences. Decreases in Northern Hemisphere summer frontal precipitation are largely driven by changes in the frequency of different front types, especially cold and stationary fronts. On the other hand, Northern Hemisphere winter exhibits some regional increases in frontal precipitation that are largely driven by changes in frontal precipitation intensity. While CONUS mean and extreme precipitation generally increase during all seasons in these climate change simulations, the likelihood of frontal extreme precipitation decreases, demonstrating that extreme precipitation has seasonally varying sources and mechanisms that will continue to evolve with climate change.