A machine learning model based on convolutional and recurrent neural networks to identify waves in the solar wind

The solar wind is permeated by wavelike fluctuations in both the magnetic field magnitude and its components, which are closely linked to other solar wind structures, such as shocks, and have significant implications for its time evolution. Understanding the role of these kinetic-scale electromagnetic waves in solar wind heating and acceleration remains an open question. In recent studies, a multibranch 1D convolutional neural network model has demonstrated high specificity in classifying circularly polarized wave modes in solar wind magnetic field data, which means good accuracy for identifying true negatives and ensuring that non-wave events are correctly classified. Building on this model's success, efforts are being made to enhance its sensitivity (ability to detect true positives, or actual wave events) without compromising the achieved specificity. To this end, recurrent neural networks are being integrated into the existing convolutional framework. This integration aims to improve the model's ability to capture temporal dependencies and subtle patterns within the data, thereby increasing the detection rate of wave intervals while maintaining the high specificity previously attained. This hybrid approach is expected to provide a more balanced performance, making it a powerful tool for the long-term cataloging and analysis of solar wind phenomena.