[Paper Review] Machine learning for predicting the Bz magnetic field component from upstream in situ observations of solar coronal mass ejections

Martin Reiß, Christian Möstl+7 more|arXiv (Cornell University)|Aug 9, 2021

Solar and Space Plasma Dynamics30 references4 citations

TL;DR

This study develops a machine learning model that predicts the minimum Bz component in interplanetary coronal mass ejections (ICMEs) using upstream in situ measurements from the sheath and first 4 hours of the magnetic obstacle. Trained on 348 ICMEs from Wind, STEREO-A, and STEREO-B, the model achieves a mean absolute error of 3.12 nT and a Pearson correlation coefficient of 0.71, demonstrating strong predictive capability for ICMEs with clear flux rope structures.

ABSTRACT

Predicting the Bz magnetic field embedded within ICMEs, also known as the Bz problem, is a key challenge in space weather forecasting. We study the hypothesis that upstream in situ measurements of the sheath region and the first few hours of the magnetic obstacle provide sufficient information for predicting the downstream Bz component. To do so, we develop a predictive tool based on machine learning that is trained and tested on 348 ICMEs from Wind, STEREO-A, and STEREO-B measurements. We train the machine learning models to predict the minimum value of the Bz component and the maximum value of the total magnetic field Bz in the magnetic obstacle. To validate the tool, we let the ICMEs sweep over the spacecraft and assess how continually feeding in situ measurements into the tool improves the Bz prediction. Because the application of the tool in operations needs an automated detection of ICMEs, we implement an existing automated ICME detection algorithm and test its robustness for the time intervals under scrutiny. We find that the predictive tool can predict the minimum value of the Bz component in the magnetic obstacle with a mean absolute error of 3.12 nT and a Pearson correlation coefficient of 0.71 when the sheath region and the first 4 hours of the magnetic obstacle are observed. While the underlying hypothesis is unlikely to solve the Bz problem, the tool shows promise for ICMEs that have a recognizable magnetic flux rope signature. Transitioning the tool to operations could lead to improved space weather forecasting.

Motivation & Objective

• Address the critical challenge of forecasting the Bz magnetic field component in ICMEs, known as the 'Bz problem' in space weather.
• Investigate whether upstream in situ measurements of the sheath and early magnetic obstacle contain sufficient information to predict downstream Bz.
• Develop a machine learning-based predictive tool for operational space weather forecasting.
• Validate the model’s performance in real-time, incremental prediction mode using ICMEs sweeping over spacecraft.

Proposed method

• Train supervised machine learning models on 348 ICME events from Wind, STEREO-A, and STEREO-B in situ measurements.
• Use input features from the sheath region and the first 4 hours of the magnetic obstacle to predict the minimum Bz and maximum total magnetic field Bt in the ICME core.
• Apply a real-time, incremental prediction framework where new measurements are fed sequentially to improve forecast accuracy over time.
• Evaluate model performance using mean absolute error (MAE) and Pearson correlation coefficient (PCC) against observed values.
• Utilize the ICMECATv2.0 catalog and in situ solar wind data from NASA's SPDF and STEREO-SSC data repositories.
• Make source code, data, and catalog publicly available via GitHub and figshare for reproducibility and community use.

Experimental results

Research questions

• RQ1 Can upstream in situ measurements of the sheath and early magnetic obstacle reliably predict the minimum Bz component in ICMEs?
• RQ2 How accurate is a machine learning model in forecasting Bz when trained on multi-spacecraft in situ observations of ICMEs?
• RQ3 Does the predictive performance improve with continuous feeding of new in situ measurements during ICME passage?
• RQ4 To what extent does the model’s accuracy depend on the presence of a well-defined magnetic flux rope structure in ICMEs?
• RQ5 Can this approach be operationalized for future space weather forecasting systems?

Key findings

• The machine learning model predicts the minimum Bz component in ICMEs with a mean absolute error (MAE) of 3.12 nT when trained on sheath and first 4 hours of the magnetic obstacle.
• The model achieves a Pearson correlation coefficient (PCC) of 0.71 between predicted and observed minimum Bz values, indicating strong predictive agreement.
• For the maximum total magnetic field Bt in the ICME core, the model achieves an MAE of 2.23 nT and a PCC of 0.91, showing even higher accuracy.
• The predictive performance improves incrementally as new in situ measurements are fed into the model during ICME passage, supporting real-time forecasting potential.
• The model shows particular promise for ICMEs with a clear magnetic flux rope signature, suggesting structural coherence enhances prediction reliability.
• The model’s framework is transferable to future missions at 1 AU and could be enhanced with semi-empirical flux rope models and uncertainty quantification.