Using Explainable Artificial Intelligence to Study Rapid Updraft Strengthening in Supercells

The low-level updraft has a significant effect on the hazards of supercells including tornadogenesis. The low-level updraft allows a supercell to stretch near-surface vorticity, providing the mechanism for tornadogenesis. By understanding the precursors to strengthening of the low-level updraft, we can understand the precursors to tornadogenesis. This has been investigated in the past using analysis of numerical simulations of supercells such as Cloud Model 1 (CM1) which allows for high resolution simulations of thunderstorms. Advances in computing have made simulating supercells in CM1 relatively inexpensive at coarser resolutions. Additionally, these advances have also allowed for the development of machine learning models which can process large volumes of data to make predictions without the complete understanding of physical processes necessary for physics-based simulation. One challenge of using machine learning for modeling complex processes has been the “black-box” characteristic of machine learning models. The field of explainable artificial intelligence (XAI) has allowed for better understanding of the patterns that machine learning models use to make predictions. By combining these approaches into a novel combined approach, we investigated the precursors to rapid low-level updraft strengthening in supercells. The first step in this approach is simulating numerous supercells in CM1 and subsetting this data around the maximum low-level updraft. A Long-Short Term Memory (LSTM) Convolutional Neural Network (CNN) was then trained on the subsetted dataset to categorize short-term changes in maximum low-level updraft strength into weakening, neutral, and strengthening. The model inputted output volumes of data from twenty-nine variables output by CM1 or derived from the CM1 simulations. XAI techniques were then applied to this LSTM CNN model to investigate the precursors to rapid low-level updraft strengthening. Two sets of techniques were used: gradient-based XAI techniques and optimal input. Gradient-based techniques use the relationship between a prediction and the input variables to trace the impacts of a part of the input on the output category. Optimal input uses back-propagation through the model to iteratively update an input until the input leads to high confidence in the input leading to the desired output. This method was applied to each of the categories and compared to determine what features of the supercells differed between the categories.