Using Machine Learning to Uncover Deep Convective Cloud Processes with TRACER Observations

Abstract

Deep convective clouds play crucial roles in our climate by contributing to Earth’s energy budget, water cycle, and vertical transport of winds, heat, moisture, and aerosol particles. Thus, it is important to accurately represent them in computer models of all scales. However, the immense complexity of processes in these clouds coupled with a lack of robust observations within them impede understanding and produce significant uncertainties in the ways our models represent them. Detailed, direct observations are critically needed to improve our understanding of these processes and to provide guidance for representing these processes in models. Fortunately, the Tracking Aerosol Convection Interactions Experiment (TRACER) field campaign, which took place in Houston, Texas, provided such much-needed observational datasets tailored for convective cloud processes. For the first time, a radar deployed as part of this campaign operated in a special storm-tracking mode that allowed it to collect thousands of high-resolution scans of deep convective clouds. This novel dataset is paired with high-resolution measurements of the environment in the vicinity these clouds collected from frequent weather balloon launches and diverse surface and profiling instruments. We will use this unprecedented quantity and quality of information to elucidate the clouds’ mixed-phase processes – i.e., those involving liquid and ice particles or hydrometeors – across cloud lifecycles and environments. By doing so, we will address the following science questions:

What are the typical hydrometeor and wind structures of convective cloud mixed-phase regions? How are these structures related to the storm’s intensity and parent environment? What are these structures’ typical growth/decay pathways and timescales?

Doing so will achieve the project’s main objectives:

• Objective 1: Characterize the major radar signals in mixed-phase regions, quantify their relative importance (in terms of explaining the variance) across the entire dataset, and provide interpretations of the microphysical and kinematic processes they represent.
• Objective 2: Quantify how these major structures vary across environments sampled during the field campaign, and determine their correlation with radar-based measures of storm intensity.
• Objective 3: Quantify the frequency content associated with these structures, their growth rates, how their emergence is related to radar-based measures of storm intensity, and their timescales.

The sheer volume and complexity of the novel radar dataset necessitates new tools and a “big data” analysis approach. To tackle this challenge, we have developed radar data preprocessing software that allows for extraction and recentering of isolated convective cloud “snapshots.” We will analyze these snapshots using high-performance machine learning methods, in which linear algebra techniques allow efficient coherent pattern extraction from large datasets, and analyses of cloud evolution. These tools will facilitate comparison of the clouds’ coherent structures to individual scans to quantify when they emerge or disappear throughout the cloud’s lifetime. We will utilize modal analyses techniques including proper orthogonal decomposition, dynamic mode decomposition, and spectral proper orthogonal decomposition on matrices constructed from the processed radar data.

Our research will lead to a robust understanding of convective mixed-phase processes and quantification of their influence on observed cross-environment and cross-lifetime variability, providing useful benchmarks for cloud process representation in computer models ranging from small-scale ones that resolve individual convective clouds through large-scale ones that simulate the entire earth system. Improved understanding of how environments modulate storm intensity and mixed-phase region properties will aid in the development of representations of convective clouds and their processes in these models. Quantifying the spectral content of cells’ coherent structures provides critical information on typical timescales and temporal relationships to other structures, painting a complete picture of convective processes through cloud lifetimes. Our research introduces powerful machine learning tools with broad applicability to the climate and atmospheric science community. Lessons learned from detailed analyses of the unprecedented radar dataset will lead to concrete recommendations for future implementations of these specialized radar scanning strategies. The cross-disciplinary nature of the proposed research and team serves as a bridge between two communities with complementary skills and expertise, fostering collaborations and new science.