Opening the Black Box of the Radiation Belt Machine Learning Model

Abstract

The Earth's radiation belts consist of energetic charged particles trapped by the geomagnetic field into two regions, a relatively stable inner zone, and a more dynamic outer zone. These particles range in energy from tens of keV to multiple MeV (e.g.,

Dependence of Electron Flux Dropouts in the Earth's Outer Radiation Belt on Energy and Driving Parameters During Geomagnetic Storms

“…Although we separately analyze the dependence of dropouts on various driving parameters, including IMF B z , P SW , SYM‐H, and AE, it is still difficult to differentiate the distinct role of different driving parameters in causing electron flux dropouts since these parameters are intrinsically correlated with each other. Future study can employ machine learning (ML) technique, such as explainable ML technique, to quantitatively analyze the attribution of various driving parameters on producing dropouts and potentially unravel the underlying physical processes (Ma et al., 2023). In addition, although the current study specifically focuses on investigation of dropouts during storm events, understanding the statistical properties of the non‐storm time dropouts are also significant and interesting, which will be left to the future study.…”

Machine Learning Interpretability of Outer Radiation Belt Enhancement and Depletion Events

“…In the following study, D. Ma et al. (2023) implemented the SHAP method to provide insight into the workings of the ML model using 2‐hr average values of AL, SYM‐H, Psw, and solar wind speed Vsw as inputs. For each input data point $$\vec{x}$$ to be explained, the sum of the SHAP values corresponds to the difference between the model prediction and the average prediction of the model for only background samples:$$\sum {\phi }_{i}=f\left(\vec{x}\right)-E\left(f\left(\vec{x}\right)\right)$$ where ϕ i is the SHAP value of the i ‐th feature.…”

“…This model has demonstrated remarkable accuracy when tested with out-of-sample data, giving R 2 ∼ 0.78-0.92 (D. Ma et al, 2022). In the following study, D. Ma et al (2023) implemented the SHAP method to provide insight into the workings of the ML model using 2-hr average values of AL, SYM-H, Psw, and solar wind speed Vsw as inputs. For each input data point 𝐴𝐴 𝐴 𝐴𝐴 to be explained, the sum of the SHAP values corresponds to the difference between the model prediction and the average prediction of the model for only background samples:…”

“…The model is driven by geomagnetic indices and solar wind parameters and successfully captures electron dynamics over long and short timescales for different energies. Building on this, D. Ma et al (2023) further examined what's inside the "Black Box" of the machine learning (ML) model with a state-of-art ML model explanation method called SHAP (Shapley Additive exPlanations, (Lundberg & Lee, 2017)). The proposed framework demonstrates that the feature importance of different input features deduced from the ML model aligns well with our existing physical understanding of both storm time and non-storm time enhancement events.…”

Influences of Solar Wind Parameters on Energetic Electron Fluxes at Geosynchronous Orbit Revealed by the Deep SHAP Method

“…colorado.edu/space_weather/xlf3/xlf3.html). In recent years, empirical models are progressively playing a more pivotal role in forecasting energetic electron fluxes at GEO based on different methods including linear prediction filters (Baker et al, 1990;Rigler et al, 2004), multiple regression (Simms et al, 2014(Simms et al, , 2018, NARMAX (Boynton et al, 2013(Boynton et al, , 2015, and artificial neural networks (ANN) (Chu et al, 2021;Koons & Gorney, 1991;D. Ma et al, 2023;Shin et al, 2016;Wang et al, 2023;Zhang et al, 2020).…”