Identifying lightning processes in ERA5 soundings with deep learning

Abstract. Atmospheric environments favorable for lightning and convection are commonly represented by proxies or parameterizations based on expert knowledge such as convective available potential energy (CAPE), wind shear, charge separation, or combinations thereof. Recent developments in the field of high-resolution reanalyses, accurate lightning observations, machine learning (ML), and explainable artificial intelligence (XAI) open possibilities for identifying tailored proxies without prior expert knowledge. This study utilizes a deep neural network trained to match temporally and vertically well-resolved ERA5 soundings of cloud physics, mass-field, and wind-field variables with lightning observations from the Austrian Lightning Information & Detection System (ALDIS). The ML model only receives the raw model atmosphere data as inputs, without incorporating any expert parameters or proxies derived from the model levels. Using and adapting appropriate XAI methods, it is then demonstrated how the inner workings of this well-performing deep learning model can be uncovered to identify physically meaningful patterns within the ERA5 soundings that describe lightning processes. The ERA5 parameters are taken on model levels beyond the tropopause, forming an input layer of approx. 670 features, and the lightning data are transformed to a binary target variable labeling the spatio-temporal ERA5 grid cells as cells with lightning activity and cells without lightning activity. Scaled Shapley additive explanations (SHAP) values are introduced to highlight the atmospheric processes learned by the neural network and show that the model identifies cloud ice and snow content in the upper troposphere and mid-troposphere as very relevant features. As these patterns correspond to the separation of charge in thunderstorm clouds, the deep learning model can serve as a physically meaningful description of lightning. The scaled SHAP values also reveal that, depending on the location, the model additionally learns to correctly classify cells with lightning activity by exploiting mass-field or wind-field variables. This approach also showcases how XAI can be used to accelerate knowledge discovery in areas where expert knowledge is still scarce.