TESS2022.pdf

Page 1 of 31

Solar Flare Forecasting with Machine Learning and Novel

Statistical Models

Yang Chen (Statistics and MIDAS)

University of Michigan, Ann Arbor

TESS, 2022

(University of Michigan) Solar Flare Forecasting TESS, 2022 1 / 25

Page 2 of 31

Outline

1 Introduction

2 Opportunities and Challenges

3 LSTM Result Demonstration

4 Mixed LSTM Regression Results Demonstration

5 Interpretable Feature Construction

6 Mixed Modeling with Multi-type Data

7 Ongoing Work

(University of Michigan) Solar Flare Forecasting TESS, 2022 2 / 25

Page 3 of 31

Introduction

1 Introduction

2 Opportunities and Challenges

3 LSTM Result Demonstration

4 Mixed LSTM Regression Results Demonstration

5 Interpretable Feature Construction

6 Mixed Modeling with Multi-type Data

7 Ongoing Work

(University of Michigan) Solar Flare Forecasting TESS, 2022 3 / 25

Page 4 of 31

Introduction

Our Past and Ongoing Work

Prediction models for flare classification1

Solar flare intensity prediction2

Understanding driving forces/features of flare eruption3

Flare classification with extended, non-calibrated, datasets4

Interpretable HMI and AIA integrated model (2022+)

Operational flare forecasting – from statistics perspective (2022+)

Probabilistic forecasting with quantified uncertainty (2022+)

1Yang Chen et al. “Identifying solar flare precursors using time series of SDO/HMI images and SHARP parameters”. In:

Space Weather 17.10 (2019), pp. 1404–1426, Xiantong Wang et al. “Predicting solar flares with machine learning: Investigating

solar cycle dependence”. In: The Astrophysical Journal 895.1 (2020), p. 3.

2Zhenbang Jiao et al. “Solar flare intensity prediction with machine learning models”. In: Space Weather 18.7 (2020),

e2020SW002440.

3Hu Sun, Ward Manchester IV, and Yang Chen. “Improved and Interpretable Solar Flare Predictions With Spatial and

Topological Features of the Polarity Inversion Line Masked Magnetograms”. In: Space Weather 19.12 (2021),

e2021SW002837, Zeyu Sun et al. “Predicting Solar Flares Using CNN and LSTM on Two Solar Cycles of Active Region Data”.

In: The Astrophysical Journal 931.2 (2022), p. 163, Spiridon Kasapis et al. “Interpretable machine learning to forecast SEP

events for solar cycle 23”. In: Space Weather 20.2 (2022), e2021SW002842.

4Zeyu Sun et al. “Predicting Solar Flares Using CNN and LSTM on Two Solar Cycles of Active Region Data”. In: The

Astrophysical Journal 931.2 (2022), p. 163.

(University of Michigan) Solar Flare Forecasting TESS, 2022 4 / 25

Page 5 of 31

Opportunities and Challenges

1 Introduction

2 Opportunities and Challenges

3 LSTM Result Demonstration

4 Mixed LSTM Regression Results Demonstration

5 Interpretable Feature Construction

6 Mixed Modeling with Multi-type Data

7 Ongoing Work

(University of Michigan) Solar Flare Forecasting TESS, 2022 5 / 25

Page 7 of 31

Opportunities and Challenges

Solar Flare Predictions: Opportunities and Challenges

Opportunities: For Both Solar Physics and ML Communities

Attractiveness: Use ML output as input for formal analysis.

High resolution images at high cadence

GOES, HMI/SDO, AIA data sets.

Challenges: From a ML/Statistics Perspective

Rare strong events: a SMALL big data problem.

Heterogeneity across events (active regions).

High resolution images at high cadence: computational efficiency.

(University of Michigan) Solar Flare Forecasting TESS, 2022 6 / 25

Page 8 of 31

Opportunities and Challenges

Solar Flare Predictions: Opportunities and Challenges

Opportunities: For Both Solar Physics and ML Communities

Attractiveness: Use ML output as input for formal analysis.

High resolution images at high cadence

GOES, HMI/SDO, AIA data sets.

Challenges: From a ML/Statistics Perspective

Rare strong events: a SMALL big data problem.

Heterogeneity across events (active regions).

High resolution images at high cadence: computational efficiency.

(University of Michigan) Solar Flare Forecasting TESS, 2022 6 / 25

Page 10 of 31

LSTM Result Demonstration

1 Introduction

2 Opportunities and Challenges

3 LSTM Result Demonstration

4 Mixed LSTM Regression Results Demonstration

5 Interpretable Feature Construction

6 Mixed Modeling with Multi-type Data

7 Ongoing Work

(University of Michigan) Solar Flare Forecasting TESS, 2022 7 / 25

Page 11 of 31

LSTM Result Demonstration

What happens for four representative, randomly selected active regions?

2011-02-11~2011-02-14

AR 11158

C-class !ares

19.6 h M6.6 M2.2

probability

1.0

0.8

0.0

0.2

0.4

0.6

2011-03-04~2011-03-07

AR 11165

C-class !ares

24.1 h M1.5 M1.5

probability

0.0

0.2

0.4

0.6

0.8

1.0

2012-06-29~2012-07-02

AR 11513

C-class !ares

19.2 h M2.4 M2.2 M1.0 M1.6 M2.8 M1.1

0 20 40 60 80 100 120

time(hours)

0.5

0.6

0.7

0.8

0.9

1.0

probability

0 20 40 60 80 100 120

time(hours)

0.2

0.4

0.6

0.8

1.0

probability

43.5 h M2.7 M6.1 M2.3

C-class !ares

AR 11532

2012-07-26~2012-07-29

84.497

(University of Michigan) Solar Flare Forecasting TESS, 2022 8 / 25

Page 12 of 31

LSTM Result Demonstration

What happens for four representative, randomly selected active regions?

2011-02-11~2011-02-14

AR 11158

C-class !ares

19.6 h M6.6 M2.2

probability

1.0

0.8

0.0

0.2

0.4

0.6

2011-03-04~2011-03-07

AR 11165

C-class !ares

24.1 h M1.5 M1.5

probability

0.0

0.2

0.4

0.6

0.8

1.0

2012-06-29~2012-07-02

AR 11513

C-class !ares

19.2 h M2.4 M2.2 M1.0 M1.6 M2.8 M1.1

0 20 40 60 80 100 120

time(hours)

0.5

0.6

0.7

0.8

0.9

1.0

probability

0 20 40 60 80 100 120

time(hours)

0.2

0.4

0.6

0.8

1.0

probability

43.5 h M2.7 M6.1 M2.3

C-class !ares

AR 11532

2012-07-26~2012-07-29

84.497

(University of Michigan) Solar Flare Forecasting TESS, 2022 8 / 25

Page 14 of 31

Mixed LSTM Regression Results Demonstration

1 Introduction

2 Opportunities and Challenges

3 LSTM Result Demonstration

4 Mixed LSTM Regression Results Demonstration

5 Interpretable Feature Construction

6 Mixed Modeling with Multi-type Data

7 Ongoing Work

(University of Michigan) Solar Flare Forecasting TESS, 2022 10 / 25

Page 17 of 31

Interpretable Feature Construction

SHARP parameter 2D distribution instead of averaged quantities.7 PIL

mask identification by Br by Schrijver (2007).

7Hu Sun, Ward Manchester IV, and Yang Chen. “Improved and Interpretable Solar Flare Predictions With Spatial and

Topological Features of the Polarity Inversion Line Masked Magnetograms”. In: Space Weather 19.12 (2021), e2021SW002837.

(University of Michigan) Solar Flare Forecasting TESS, 2022 13 / 25

Page 22 of 31

Mixed Modeling with Multi-type Data

1 Introduction

2 Opportunities and Challenges

3 LSTM Result Demonstration

4 Mixed LSTM Regression Results Demonstration

5 Interpretable Feature Construction

6 Mixed Modeling with Multi-type Data

7 Ongoing Work

(University of Michigan) Solar Flare Forecasting TESS, 2022 18 / 25

Page 25 of 31

Mixed Modeling with Multi-type Data

Flare Prediction: Experiment Design & Results

2010-2018: 3,064 B-class flares and 586 M-class flares.

Forecasting window: 1,6,12,24 hours before peak time.

images resized to: 50 × 50, 100 × 100, 200 × 200.

Pred Time 50 × 50 100 × 100 200 × 200 SHARP only

1-hour 0.9143 0.9069 0.9201 0.9078

6-hour 0.9161 0.9116 0.9316 0.9141

12-hour 0.8953 0.8923 0.9061 0.8833

24-hour 0.9169 0.8986 0.9172 0.8898

Table 1: Area Under Curve (AUC) for Binary Flare Classification for all 4

prediction time and 3 Br

image sizes. Metrics based on four MCMC chains.

(University of Michigan) Solar Flare Forecasting TESS, 2022 21 / 25

Page 26 of 31

Ongoing Work

1 Introduction

2 Opportunities and Challenges

3 LSTM Result Demonstration

4 Mixed LSTM Regression Results Demonstration

5 Interpretable Feature Construction

6 Mixed Modeling with Multi-type Data

7 Ongoing Work

(University of Michigan) Solar Flare Forecasting TESS, 2022 22 / 25

Page 28 of 31

Ongoing Work

Interpretable HMI and AIA integrated model (2022+)

AIA database construction (Meng Jin & Yang Liu)

Integrated learning algorithm (code ready)

Operational flare forecasting – from statistics perspective (2022+)

Summary of methodology in literature

Gaps with operational forecasting

Extensive simulation experiments

Probabilistic forecasting with quantified uncertainty (2022+)

AR temporal history informed

Calibrated prediction intervals for arrival & intensity

(University of Michigan) Solar Flare Forecasting TESS, 2022 23 / 25

Page 30 of 31

Ongoing Work

References I

[1] Yang Chen et al. “Identifying solar flare precursors using time series

of SDO/HMI images and SHARP parameters”. In: Space Weather

17.10 (2019), pp. 1404–1426.

[2] Zhenbang Jiao et al. “Solar flare intensity prediction with machine

learning models”. In: Space Weather 18.7 (2020), e2020SW002440.

[3] Spiridon Kasapis et al. “Interpretable machine learning to forecast

SEP events for solar cycle 23”. In: Space Weather 20.2 (2022),

e2021SW002842.

[4] Hu Sun, Ward Manchester IV, and Yang Chen. “Improved and

Interpretable Solar Flare Predictions With Spatial and Topological

Features of the Polarity Inversion Line Masked Magnetograms”. In:

Space Weather 19.12 (2021), e2021SW002837.

(University of Michigan) Solar Flare Forecasting TESS, 2022 24 / 25

Page 31 of 31

Ongoing Work

References II

[5] Zeyu Sun et al. “Predicting Solar Flares Using CNN and LSTM on

Two Solar Cycles of Active Region Data”. In: The Astrophysical

Journal 931.2 (2022), p. 163.

[6] Xiantong Wang et al. “Predicting solar flares with machine learning:

Investigating solar cycle dependence”. In: The Astrophysical Journal

895.1 (2020), p. 3.

(University of Michigan) Solar Flare Forecasting TESS, 2022 25 / 25