A deep learning framework for analyzing cloud characteristics of aggregated convection using cloud-resolving model simulations

Abstract

This study introduces a framework to extract the high-dimensional nonlinear relationships among state variables for aggregated convection. The prototype of such a framework is developed that applies the convolutional neural network models (CNN models) to retrieve the cloud characteristics from cloud-resolving model (CRM) simulations. CNN model prediction factors are hidden in the high dimensional weighted parameters in each neural network layer. Therefore, we can dig out relevant physics processes by iterating the CNN models' training process and eliminating the features with the physics explanation we can provide at a given stage. Within a few iterations, explainable nonlinear relationships among variables can be provided. We identified that the averaged cloud water path (CWP), the maximum value of CWP in each cloud, and the cloud coverage rate are essential for identifying aggregation. Furthermore, by analyzing the encoded channels of the CNN model, we found a strong relationship between aggregation, cloud peripherals, and fractal dimensions. The results suggest that the important nonlinear cloud characteristics for identifying the aggregation can be captured with the proper adjustment and limitation of the input data to the CNN models. Our framework provides a possibility that we can explore the high dimensional relationship between the physics process with the assistance of the CNN model.