Authors:

William Henry Stevens

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Atmospheric blocking events, characterized by persistent high-pressure systems that block prevailing westerly wind patterns, can significantly impact weather patterns and lead to extreme conditions like heat waves, cold spells and droughts. However, there has yet to be any significant study identifying the precursor patterns that indicate the imminence of blocking long before it occurs. This project explores a multidisciplinary approach to identifying and understanding these precursor patterns through integrating a scientific understanding of physics and atmospheric science with the patterns identified by a Convolutional Neural Network (CNN) and Explainable Artificial Intelligence (XAI). The network is trained on simulation data from the Two-Layer Quasi- Geostrophic (QG) Model, investigating various atmospheric variables such as potential vorticity and jetstream flow, and has shown to predict the occurrence of future atmospheric blocking with 95%, 90%, and 85% accuracy at 1, 5, and 10 lead days prior to the event, respectively. The patterns used by the neural network to predict the onset of blocking were identified through employment of an XAI technique called Layer-Wise Relevance Propagation (LRP), which produces a relevancy heat-map highlighting the pixel features deemed most important to the network's predictions. These features are consistent with a physics-based understanding of precursor blocking behavior. This work confirms the existence of precursor patterns for atmospheric blocking, identifies their behavior, and proposes a reliable mechanism for using them to predict future blocking occurrences. The promising results motivate future research endeavors utilizing machine learning and physical understanding to predict atmospheric blocking long before its effects impact the world. Keywords: Machine Learning; Atmospheric Science; Atmospheric Blocking; Explainable AI; Heatwaves

Source:

Purdue University / 2024

Topics:

No topics listed

Co-authors:

William Henry Stevens