Cosmic-Ray-Constrained LSTM Model for Geomagnetic Storm Prediction

TL;DR

This paper introduces a physics-informed LSTM model that uses cosmic-ray flux data as a precursor to improve geomagnetic storm predictions, achieving better accuracy and longer warning times.

Contribution

It is the first to incorporate cosmic-ray flux modulation into an LSTM model for geomagnetic storm prediction, enhancing forecast skill with multi-source space-weather data.

Findings

Cosmic-ray data improves 48-hour forecast accuracy by up to 25.84%.

The model achieves RMSEs of 5.106 nT to 14.788 nT across 2- to 48-hour predictions.

Multi-source data integration enhances space-weather forecasting capabilities.

Abstract

Geomagnetic storms (GSTs) driven by solar wind-magnetosphere coupling can severely disrupt technological systems, motivating the need for improved prediction accuracy and longer warning times. In this study, we develop a physics-informed Long Short-Term Memory (LSTM) model that incorporates cosmic-ray flux modulation as an additional precursor signal. As coronal mass ejection (CME)-driven disturbances propagate through the heliosphere, enhanced turbulence and magnetic-field compression reduce galactic cosmic-ray (GCR) flux measured by ground-based neutron monitors (Forbush decreases), providing early information that can precede near-Earth solar-wind signatures by 1--3 days. We integrate multi-source space-weather data, spanning 1995-2020, including cosmic-ray observations, solar wind plasma parameters, interplanetary magnetic-field data, and geomagnetic indices. Based on these data, we construct a 19-dimensional feature vector that includes flux background levels, decrease-related indicators, and inter-station correlation measures as model inputs. Employing a 50-unit LSTM architecture, the proposed model achieves root-mean-square errors (RMSE) of 5.106 nT, 8.315 nT, 10.854 nT, 12.883 nT, and 14.788 nT for 2-, 6-, 12-, 24-, and 48-hour predictions, respectively. Incorporating cosmic-ray information further improves 48-hour forecast skill by up to 25.84% (from 0.178 to 0.224). These results demonstrate the value of physics-informed deep learning and cosmic-ray precursors for advancing space-weather forecasting.