ML Eco Fi
A Machine Learning Ecosystem for Filament Detection:
Classification, Localization, and Segmentation
[2022 - 2025]
This multi-year project is supported by an NSF grant award with award number #2209912, #2433781.
PROJECT OVERVIEW
Extreme space-weather events, similar to extreme terrestrial events, can have economic and collateral impacts on mankind. The energetic particles of strong Coronal Mass Ejections (CMEs) can reach the Earth in a course of 1-5 days and can cause geomagnetic storms with direct impacts on our electronic infrastructures (causing electric power outage) and the GPS system (disturbing the GPS-based positioning industries), resulting in economic damage and tragic social consequences, including substantial loss of life. The National Research Council reported that a solar superstorm could cripple the entire US power grid for months and lead to economic damage of $1-2 trillion. Large CMEs and high intensity solar flares are almost always initiated and driven by the eruption of solar filaments. Analyzing filaments can tell us the coordinates from which a potential CME would originate, and the magnetic structure of the nearby filaments, helping a CME forecast model to estimate the direction of the expulsion. Therefore, a network of six observatories distributed around the globe, called the Global Oscillation Network Group (GONG), exists to monitor the solar disk for filaments 24/7 to catch eruptions as they happen. Our proposed project enables interdisciplinary research with true Big Data capabilities.
GONG’s uninterrupted observation, with a one-minute cadence, can only be fully harvested with the help of the past decade’s fast and smart computer vision algorithms which manifested an unprecedented success thanks to the re-emergence of the Deep Neural Networks. We propose an ML Ecosystem that will provide, for the first time, the largest ever manually-annotated dataset of H-alpha images (providing chirality, bounding box, and segmentation mask for each filament), advancing the research on filaments and filament-related topics. Moreover, this ecosystem will provide the community with two essential components for ML algorithms, namely, a Chirality-aware Filament Augmentation Engine and a high-precision loss function. Additionally, we will produce a filament-detection module that will put one million annotated filaments (and their tracking information) at the community’s disposal, making it possible to test many hypotheses and build new theories. With collaboration with a team from the National Solar Observatory (NSO), we will deploy our module on NSO’s infrastructure to serve the community as a live, high-precision filament-detection module.