ML-GEM Discussion Session (10:30am-12pm on Friday)

We will share currently available ML models and discuss how to unify them into a single AI-driven modeling framework for understanding the Sun - Earth interaction from a data-driven perspective.

​

Please add maximum two slides of your ML model to this google file by 6pm Thursday, Jun 27.

​

Your slides should follow the following format:

• 1st slide : brief description of your model (e.g., data used in AI, AI technique, input and output of model)
• 2nd slide : example figures about the model output

​

Please add your model slides below

Solar Wind - Magnetosphere Interaction (SWMI)

Deep Learning model of the Northern Earth Cusp

Authors: Gonzalo Cucho-Padin, David Sibeck, Daniel da Silva

Affiliation: NASA/GSFC

DATA: 17 years of NASA OMNI & ESA’s Cluster ion flux data (CIS-CODIF, C1/3/4)

• Training dataset : 2001-2017 (Training 80% and Validation 20%). 

�Method: Residual Neural Networks (non-linear regression model, 28 residual blocks)

Inputs:  

Cluster location (XYZ in SM), Dipole tilt, and time history of solar wind velocity, density, IMF Bx, By, Bz, and geomagnetic indexes (AL,AU,SymH) at t-30min

​

Outputs:

Ion energy flux in the 31 CIS/CODIF energy channels (25eV – 35keV) �

C1, C2, C3 ,C4

Cluster mission

Observed Spectrum

Region of Interest

Deep Learning model of the Northern Earth Cusp

Authors: Gonzalo Cucho-Padin, David Sibeck, Daniel da Silva

Affiliation: NASA/GSFC

0.14 keV

1.2 keV

2 keV

Log(Ion Flux)

Y_SM [RE]

X_SM [RE]

X_SM [RE]

X_SM [RE]

Log(Ion Flux)

For a given set of solar wind parameters, geomagnetic indexes, and current dipole tilt, we can generate the 3-D structure of the ion flux cusp.

Magnetotail and Plasmasheet (MPS)

Magnetosphere - Ionosphere Coupling (MIC)

ML Particle Precipitation Model (ML-PPM)

DATA: 17 years of NASA OMNI & DMSP SSJ electron data

• Training dataset : 1987-2013 (Training 80% and Validation 20%).
• Test data set : a whole year in 2014

​

Method: Residual Neural Network

• Inputs:

MLAT, MLT, DOY, UT, F10.7, and time histories of solar wind velocity, pressure, IMF Bx, By and Bz, and geomagnetic indices (AL, AU, SymH, PC) at t-6hr, t-3hr, t-1hr, t-45min, t-30min, t-10min)

​

• Outputs:

Differential energy flux of precipitating electrons in the 19 DMSP energy channels (30eV – 30keV) in the resolution of 1 MLT x 1° MLAT and 1 minute .

Valluri Sai Gowtam (svalluri@alaska.edu)

& Hyunju Connor (hyunju.k.connor@nasa.gov)

ML-PPM Outputs

Northward

IMF Bz

Weakly southward

IMF Bz

Strongly southward

IMF Bz

Valluri Sai Gowtam (svalluri@alaska.edu)

& Hyunju Connor (hyunju.k.connor@nasa.gov)

Deep leArninG Geomagnetic pErtuRbation (DAGGER-2020) 

B.Ferdousi, V. Upendran, M. Heyns, R. Mukundan

Targets: SuperMAG

Geomagnetic Field Measurements

Examples from Upendran et. al., 2022

Predictions: Global Spherical Harmonic Basis Estimates of Geomagnetic Field

�Approved for public release; distribution is unlimited. Public Affairs release approval #-24-03977.

12

DISTRIBUTION D: Distribution authorized to Department of Defense and US DoD contractors only (Critical Technology, 8 November 2023). Refer other requests for this document to AFRL/RVBX, Kirtland AFB, NM.

Results of DAGGER++

Target

Base Model-v1

All-Experiments

Geomagnetic Indices

Stat-Regularization

Imbalanced Regression

Imbalanced Regression Correction

​

• Data density difference between storm time and quiet time
• Adjust loss function weight per timestep for storm time data

Geomagnetic Indices

​

• Improve accuracy through the inclusion of current state of geospace
• Incorporate additional features, i.e., SME, SMU/L, SMR

Ground Station Regularization

​

• Allow higher spherical harmonic orders to capture more localized signatures
• Adjust loss function weight per target by station density

�Approved for public release; distribution is unlimited. Public Affairs release approval #-24-03977.

B.Ferdousi, V. Upendran, M. Heyns, R. Mukundan

13

DISTRIBUTION D: Distribution authorized to Department of Defense and US DoD contractors only (Critical Technology, 8 November 2023). Refer other requests for this document to AFRL/RVBX, Kirtland AFB, NM.

SHEATH

DAGGER

Coronal hole & active region masking, feature generation

Solar wind regressor

Solar wind forecasts: days in advance

Magnetospheric encoder

�Approved for public release; distribution is unlimited. Public Affairs release approval #-24-03977.

B.Ferdousi, V. Upendran, M. Heyns, R. Mukundan

14

DISTRIBUTION D: Distribution authorized to Department of Defense and US DoD contractors only (Critical Technology, 8 November 2023). Refer other requests for this document to AFRL/RVBX, Kirtland AFB, NM.

SHEATH-DAGGER Pipeline — Multi-Day Ahead Forecasts

�Approved for public release; distribution is unlimited. Public Affairs release approval #-24-03977.

B.Ferdousi, V. Upendran, M. Heyns, R. Mukundan

15

DISTRIBUTION D: Distribution authorized to Department of Defense and US DoD contractors only (Critical Technology, 8 November 2023). Refer other requests for this document to AFRL/RVBX, Kirtland AFB, NM.

Residual Convolutional Networks for Geomagnetic Field Modeling

Data: 21 years ACE SWEPAM & MAG instrument data and SuperMAG Gridded Data (Kp 5 or higher storms)

​

Model: Residual Neural Network

​

Inputs: IMF |B|, B_Y, B_Z, V_SW, N_SW, T_SW, P_SW, E_SW @ t-30 to t-90

​

Output: B_N @ 2 MLAT x 1 MLT resolution from 40 to 90 MLAT (600 grid points)

Matthew Blandin (mjblandin@alaska.edu)

Hyunju K. Connor (hyunju.k.connor@nasa.gov)

Follows Gridded data product well

​

Does not quite capture highly localized station values

​

Scores are higher and error is lower than current empirical techniques

​

Improves upon recent CNN modeling effort

Convolutional Neural Network Forecasting Geomagnetic Localization

μ

σ

Inner Magnetosphere (IMAG)

• Develop a new hybrid framework: Integrate deep learning into physics-based simulations to model global electron precipitation driven by chorus and hiss
• Data of density, chorus and hiss wave amplitude, and electron flux are from Van Allen Probes from 2012-2019.
• Density and wave amplitudes are used as inputs to Full Diffusion Code to calculate diffusion coefficients.
• The QL approach of Ma et al. (2020, 2021) is used to calculate precipitating electron flux.

Modeling Global Electron Precipitation Driven by Whistler Mode Waves

Sheng Huang (hs2015@bu.edu) & Wen Li (wenli77@bu.edu) & Qianli Ma (qma@bu.edu)

Application to 17 March 2013 geomagnetic storm

Sheng Huang (hs2015@bu.edu) & Wen Li (wenli77@bu.edu) & Qianli Ma (qma@bu.edu)

Simulation results using deep-learning inputs capture the dynamics of chorus and hiss, and their associated electron precipitation on a global scale.

ML models for the inner magnetosphere

collected by Xiangning Chu (chuxiangning@gmail.com)

1. Cold plasma density (Bortnik et al., 2016; Chu et al., 2017a,b; Zhelavskaya et al., 2017, 2020; Ace et al., 2021; Huang et al., 2022; Ripoll et al., 2023).
2. Whistler-mode waves (Bortnik et al., 2018; Chu et al., 2022, 2024; Guo et al., 2022; Huang et al., 2023; Kim et al., 2015)
3. Electron fluxes (Chu et al., 2021; Ma D. et al., 2022)
4. Ring current ion fluxes (Li et al., 2023; Wang et al., 2024)
5. ML+Physics-based Fokker-Planck (Bortnik et al., 2018; Ma Q et al., 2018; Huang et al., 2023)
6. Interpretable ML (Ma D. et al., 2023a,b).
7. GIC model (Cao et al., 2023)
8. Uncertainty of models (Camporeale et al., 2019; Chu et al., 2023;2024)
9. Magnetic field model on Earth and Mars (Cao et al., 2024)

ML models for the inner magnetosphere

collected by Xiangning Chu (chuxiangning@gmail.com)

Inputs:

Time series of geomagnetic indices and upstream solar wind conditions

Satellite location

Output/target:

cold plasma density (log)

electron and ion fluxes (log)

upper hybrid resonance frequency (bin)

plasmapause location (linear)

wave amplitude (log)

Equivalent ionospheric current (linear)

Error of models (log or linear)

Contributions from inputs (for interpretable ML)

​

​

ML models for the inner magnetosphere

collected by Xiangning Chu (chuxiangning@gmail.com)

• Cold plasma density (Bortnik et al., 2016; Chu et al., 2017a,b; Zhelavskaya et al., 2017, 2020; Ace et al., 2021; Huang et al., 2022; Ripoll et al., 2023).
• Plasmapause location (Guo et al., 2021)
• Whistler-mode waves (Bortnik et al., 2018; Chu et al., 2022, 2024; Guo et al., 2022; Huang et al., 2023; Kim et al., 2015)
• Electron fluxes (Chu et al., 2021; Ma et al., 2022)
• Ring current ion fluxes (Li et al., 2023; Wang et al., 2024)
• ML+Physics-based Fokker-Planck (Bortnik et al., 2018; Ma et al., 2018; Huang et al., 2023)
• Interpretable ML (Ma et al., 2023a,b).
• GIC model (Cao et al., 2023)
• Uncertainty of models (Camporeale et al., 2019; Chu et al., 2023;2024)

ML models for the inner magnetosphere

collected by Xiangning Chu (chuxiangning@gmail.com)

• Cold plasma density (Bortnik et al., 2016; Chu et al., 2017a,b; Zhelavskaya et al., 2017, 2020; Ace et al., 2021; Huang et al., 2022; Ripoll et al., 2023).
• Plasmapause location (Guo et al., 2021)
• Whistler-mode waves (Bortnik et al., 2018; Chu et al., 2022, 2024; Guo et al., 2022; Huang et al., 2023; Kim et al., 2015)
• Electron fluxes (Chu et al., 2021; Ma et al., 2022)
• Ring current ion fluxes (Li et al., 2023; Wang et al., 2024)
• ML+Physics-based Fokker-Planck (Bortnik et al., 2018; Ma et al., 2018; Huang et al., 2023)
• Interpretable ML (Ma et al., 2023a,b).
• GIC model (Cao et al., 2023)
• Uncertainty of models (Camporeale et al., 2019; Chu et al., 2023;2024)

ML models for the inner magnetosphere

collected by Xiangning Chu (chuxiangning@gmail.com)

• Cold plasma density (Bortnik et al., 2016; Chu et al., 2017a,b; Zhelavskaya et al., 2017, 2020; Ace et al., 2021; Huang et al., 2022; Ripoll et al., 2023).
• Plasmapause location (Guo et al., 2021)
• Whistler-mode waves (Bortnik et al., 2018; Chu et al., 2022, 2024; Guo et al., 2022; Huang et al., 2023; Kim et al., 2015)
• Electron fluxes (Chu et al., 2021; Ma et al., 2022)
• Ring current ion fluxes (Li et al., 2023; Wang et al., 2024)
• ML+Physics-based Fokker-Planck (Bortnik et al., 2018; Ma et al., 2018; Huang et al., 2023)
• Interpretable ML (Ma et al., 2023a,b).
• GIC model (Cao et al., 2023)
• Uncertainty of models (Camporeale et al., 2019; Chu et al., 2023;2024)

ML models for the inner magnetosphere

collected by Xiangning Chu (chuxiangning@gmail.com)

• Cold plasma density (Bortnik et al., 2016; Chu et al., 2017a,b; Zhelavskaya et al., 2017, 2020; Ace et al., 2021; Huang et al., 2022; Ripoll et al., 2023).
• Plasmapause location (Guo et al., 2021)
• Whistler-mode waves (Bortnik et al., 2018; Chu et al., 2022, 2024; Guo et al., 2022; Huang et al., 2023; Kim et al., 2015)
• Electron fluxes (Chu et al., 2021; Ma et al., 2022)
• Ring current ion fluxes (Li et al., 2023; Wang et al., 2024)
• ML+Physics-based Fokker-Planck (Bortnik et al., 2018; Ma et al., 2018; Huang et al., 2023)
• Interpretable ML (Ma et al., 2023a,b).
• GIC model (Cao et al., 2023)
• Uncertainty of models (Camporeale et al., 2019; Chu et al., 2023;2024)

ML models for the inner magnetosphere

collected by Xiangning Chu (chuxiangning@gmail.com)

• Cold plasma density (Bortnik et al., 2016; Chu et al., 2017a,b; Zhelavskaya et al., 2017, 2020; Ace et al., 2021; Huang et al., 2022; Ripoll et al., 2023).
• Plasmapause location (Guo et al., 2021)
• Whistler-mode waves (Bortnik et al., 2018; Chu et al., 2022, 2024; Guo et al., 2022; Huang et al., 2023; Kim et al., 2015)
• Electron fluxes (Chu et al., 2021; Ma et al., 2022)
• Ring current ion fluxes (Li et al., 2023; Wang et al., 2024)
• ML+Physics-based Fokker-Planck (Bortnik et al., 2018; Ma et al., 2018; Huang et al., 2023)
• Interpretable ML (Ma et al., 2023a,b).
• GIC model (Cao et al., 2023)
• Uncertainty of models (Camporeale et al., 2019; Chu et al., 2023;2024)

ML models for the inner magnetosphere

collected by Xiangning Chu (chuxiangning@gmail.com)

• Cold plasma density (Bortnik et al., 2016; Chu et al., 2017a,b; Zhelavskaya et al., 2017, 2020; Ace et al., 2021; Huang et al., 2022; Ripoll et al., 2023).
• Plasmapause location (Guo et al., 2021)
• Whistler-mode waves (Bortnik et al., 2018; Chu et al., 2022, 2024; Guo et al., 2022; Huang et al., 2023; Kim et al., 2015)
• Electron fluxes (Chu et al., 2021; Ma et al., 2022)
• Ring current ion fluxes (Li et al., 2023; Wang et al., 2024)
• ML+Physics-based Fokker-Planck (Bortnik et al., 2018; Ma et al., 2018; Huang et al., 2023)
• Interpretable ML (Ma et al., 2023a,b).
• GIC model (Cao et al., 2023)
• Uncertainty of models (Camporeale et al., 2019; Chu et al., 2023;2024)

ML models for the inner magnetosphere

collected by Xiangning Chu (chuxiangning@gmail.com)

• Cold plasma density (Bortnik et al., 2016; Chu et al., 2017a,b; Zhelavskaya et al., 2017, 2020; Ace et al., 2021; Huang et al., 2022; Ripoll et al., 2023).
• Plasmapause location (Guo et al., 2021)
• Whistler-mode waves (Bortnik et al., 2018; Chu et al., 2022, 2024; Guo et al., 2022; Huang et al., 2023; Kim et al., 2015)
• Electron fluxes (Chu et al., 2021; Ma et al., 2022)
• Ring current ion fluxes (Li et al., 2023; Wang et al., 2024)
• ML+Physics-based Fokker-Planck (Bortnik et al., 2018; Ma et al., 2018; Huang et al., 2023)
• Interpretable ML (Ma et al., 2023a,b).
• GIC model (Cao et al., 2023)
• Uncertainty of models (Camporeale et al., 2019; Chu et al., 2023;2024)

ML models for the inner magnetosphere

collected by Xiangning Chu (chuxiangning@gmail.com)

• Cold plasma density (Bortnik et al., 2016; Chu et al., 2017a,b; Zhelavskaya et al., 2017, 2020; Ace et al., 2021; Huang et al., 2022; Ripoll et al., 2023).
• Plasmapause location (Guo et al., 2021)
• Whistler-mode waves (Bortnik et al., 2018; Chu et al., 2022, 2024; Guo et al., 2022; Huang et al., 2023; Kim et al., 2015)
• Electron fluxes (Chu et al., 2021; Ma et al., 2022)
• Ring current ion fluxes (Li et al., 2023; Wang et al., 2024)
• ML+Physics-based Fokker-Planck (Bortnik et al., 2018; Ma et al., 2018; Huang et al., 2023)
• Interpretable ML (Ma et al., 2023a,b).
• GIC model (Cao et al., 2023)
• Uncertainty of models (Camporeale et al., 2019; Chu et al., 2023;2024)

ML models for the inner magnetosphere

collected by Xiangning Chu (chuxiangning@gmail.com)

• Cold plasma density (Bortnik et al., 2016; Chu et al., 2017a,b; Zhelavskaya et al., 2017, 2020; Ace et al., 2021; Huang et al., 2022; Ripoll et al., 2023).
• Plasmapause location (Guo et al., 2021)
• Whistler-mode waves (Bortnik et al., 2018; Chu et al., 2022, 2024; Guo et al., 2022; Huang et al., 2023; Kim et al., 2015)
• Electron fluxes (Chu et al., 2021; Ma et al., 2022)
• Ring current ion fluxes (Li et al., 2023; Wang et al., 2024)
• ML+Physics-based Fokker-Planck (Bortnik et al., 2018; Ma et al., 2018; Huang et al., 2023)
• Interpretable ML (Ma et al., 2023a,b).
• GIC model (Cao et al., 2023)
• Uncertainty of models (Camporeale et al., 2019; Chu et al., 2023;2024)

Lessons learned:

1. Models are good, but physics are more important
2. Data imbalance and imbalanced regression
1. Imbalanced regression is less investigated, even in ML field (not imbalanced classification)
2. The existing imbalanced regression methods do not work on my database. Database-specific weighting methods are required.
3. Model trained on ‘interesting events’ (e.g., storm time only) is essentially weighting data samples. It may achieve certain goals but I’ve found many glitches in the past studies that are not shown in the paper.
3. Integration of models are important
• We have coupled many models together to achieve more accurate and physics-reliable models
• Collaborations are important.

ML models for the inner magnetosphere

collected by Xiangning Chu (chuxiangning@gmail.com)

Planets and Heliosphere

• Mars does not have an upstream solar wind monitor; fundamentally limits studies in: Mars-solar wind interactions, planetary comparisons, and solar wind propagation.

​

• We used MAVEN data [Halekas+2017] with Gaussian process regression to create a continuous estimate of the solar wind (B_x, B_y, B_z, |B| v_x, v_y, v_z, |v|, T_p, n_p) with uncertainty.

​

A Virtual Solar Wind Monitor (vSWIM) at Mars

A. R. (Abby) Azari (Data Science Post Doc Fellow at UBC azari@eoas.ubc.ca), E. Abrahams, F. Sapienza, J. Halekas, J. Biersteker,

D. L. Mitchell, F. Pérez, M. Marquette, M. J, Rutala, C. F. Bowers, C. M. Jackman, & S. M. Curry

​

This improves on previous studies [see Hurley+2018, Ruhunusiri+2019, Dong+2019] to enable highly accurate estimation of the solar wind upstream within 2 days (66% of the time, R²≥0.95) of a true solar wind measurement and reasonable estimation within 10 days (95% of the dataset, R²≥0.62).

​

A Virtual Solar Wind Monitor (vSWIM) at Mars

A. R. (Abby) Azari (Data Science Post Doc Fellow at UBC azari@eoas.ubc.ca), E. Abrahams, F. Sapienza, J. Halekas, J. Biersteker,

D. L. Mitchell, F. Pérez, M. Marquette, M. J, Rutala, C. F. Bowers, C. M. Jackman, & S. M. Curry

ML-GEM Tutorial On LSTM