CLARE: Classification-based Regression for Electron Temperature Prediction

The paper introduces CLARE, a classification-based regression machine learning model that significantly improves electron temperature prediction accuracy and provides uncertainty estimates by transforming the continuous output space into discrete intervals using AKEBONO satellite data.

The Gist

The "Weather Forecast" for Earth's Invisible Bubble

Imagine Earth is wrapped in a giant, invisible bubble made of super-cold gas (plasma). Scientists call this the plasmasphere. It's like a protective shield for our satellites, but it's also a chaotic place where the temperature of tiny particles (electrons) can swing wildly.

Knowing the temperature of these electrons is crucial. If we get it wrong, our satellites might malfunction, GPS could glitch, and astronauts could be exposed to dangerous radiation. But for decades, predicting this temperature has been like trying to guess the exact temperature of a pot of soup by looking at the steam from across the room. It's hard, the data is messy, and the old "recipes" (physics models) often miss the mark.

Enter CLARE (Classification-based Regression for Electron Temperature), a new artificial intelligence (AI) model created by a team of scientists that acts like a super-smart weather forecaster for this invisible bubble.

Here is how CLARE works, broken down into simple ideas:

1. The Problem: Guessing a Number vs. Guessing a Range

Imagine you are playing a game where you have to guess the exact temperature of a room.

• The Old Way (Continuous Regression): You try to guess the exact number, like "It's 72.43 degrees." If you guess 72.44, you are technically wrong. This is hard because the data is noisy, like trying to hear a whisper in a hurricane.
• The CLARE Way (Classification): Instead of guessing the exact number, CLARE divides the temperature scale into 150 buckets (bins). It asks, "Is the temperature in the 70–71 degree bucket? The 71–72 degree bucket?" It picks the most likely bucket and then gives you the middle of that bucket as the answer.

The Analogy: Think of it like a dartboard.

• The Old Way tries to hit the exact bullseye (the perfect number). If you miss by a millimeter, you lose.
• CLARE divides the board into large colored zones. It just needs to land in the right colored zone. Once it lands there, it says, "I'm pretty sure it's the center of this zone." This makes it much more accurate and less likely to be thrown off by a tiny bit of wind (noise).

2. The Secret Sauce: Confidence Levels

Because CLARE looks at all 150 buckets, it can tell you how sure it is.

• If it thinks 90% of the probability is in one specific bucket, it's very confident.
• If the probability is spread out across 20 different buckets, it's saying, "I'm not sure, the conditions are chaotic."

This is like a human weather forecaster saying, "It will rain" vs. "It might rain, but I'm not sure." This built-in "confidence meter" is a huge bonus that older models didn't have.

3. The Results: A Giant Leap Forward

The scientists tested CLARE against the old standard models (which are like the "weather maps" from 20 years ago).

• On calm days: The old models got the temperature right within a 10% margin only about 13% of the time. CLARE got it right 69% of the time. That's a massive improvement!
• During Solar Storms: Solar storms are like hurricanes for space. They are rare and chaotic. The old models almost completely failed during these times (getting it right only 5% of the time). CLARE did much better, getting it right 46% of the time.
  • Why not 100%? Solar storms are so rare in the training data (less than 1% of the time) that the AI hasn't seen enough of them to become a master at predicting them yet. But it's still a huge step up.

4. Why This Matters

Think of the plasmasphere as the "ocean" that our satellites swim in.

• Before CLARE: We were navigating the ocean with a blurry, outdated map.
• With CLARE: We now have a high-definition, real-time sonar that tells us the water temperature and how confident it is in its reading.

This helps engineers design better satellites, keeps GPS accurate, and protects astronauts. It proves that even with messy, incomplete data, modern AI can learn the hidden patterns of space physics better than traditional math equations ever could.

In a nutshell: CLARE is a new AI that predicts space weather by grouping temperatures into "buckets" instead of guessing exact numbers. It's more accurate, knows when it's unsure, and is already outperforming the best scientific models we had before.

Technical Summary

1. Problem Statement

The Challenge: Electron temperature ($T_e$) is a critical parameter governing space weather dynamics in Earth's plasmasphere (1000–8000 km altitude). However, predicting $T_e$ has historically been underexplored in machine learning (ML) literature compared to electron density ($N_e$), primarily due to data scarcity and the complexity of the physical system.
Limitations of Existing Approaches:

• Physics-based models: While grounded in first principles, they are computationally intensive and struggle to capture the full complexity of interacting drivers (e.g., solar activity, geomagnetic disturbances).
• Traditional Empirical Models: Models like Titheridge (1998) and IRI-2020 rely on averaging long-term observations, often failing to capture non-linear, real-time dynamics, especially during geomagnetic storms.
• Standard ML Regression: Treating $T_e$ prediction as a standard continuous regression task often suffers from high variance, sensitivity to noise, and a lack of built-in uncertainty quantification.

2. Methodology: The CLARE Model

The authors introduce CLARE (Classification-based Regression for Electron Temperature), a novel deep learning architecture designed to overcome the limitations of standard regression.

Core Innovation: Classification-based Regression
Instead of predicting a single continuous scalar value, CLARE discretizes the continuous $T_e$ output space into 150 discrete bins (each 100 K wide, covering 0–15,000 K).

• Architecture: A 6-layer feedforward neural network with 84 million parameters. It utilizes a stack of 5 feedforward blocks (Linear $\to$ LayerNorm $\to$ ReLU $\to$ Dropout) followed by a final projection layer.
• Output Mechanism: The final layer outputs 150 neurons representing a probability distribution over the bins.
• Training Objective: The model is trained using Cross-Entropy Loss, treating the target as a classification problem (identifying the correct bin index) rather than Mean Squared Error (MSE) regression.
• Inference: The final $T_e$ value is derived by applying argmax to the predicted probabilities and shifting the result to the center of the selected bin ($T_e = k \cdot \Delta T + 0.5\Delta T$).

Data Sources

• Target Data: In-situ thermal electron distribution (TED) measurements from the AKEBONO (EXOS-D) satellite (1990–2001), filtered for altitudes between 1000 km and 8000 km.
• Input Features:
  • Spatial: Location data (L-shell, Magnetic Latitude, Geomagnetic Local Time).
  • Solar/Geomagnetic Indices: SYM-H, AL, F10.7, and Kp indices from the NASA OMNI dataset.

Uncertainty Quantification
By outputting a probability distribution over bins, CLARE provides intrinsic uncertainty estimation. Sharp distributions indicate high confidence, while broad distributions indicate high uncertainty (e.g., during chaotic storm events).

3. Key Contributions

1. First ML Model for Plasmaspheric $T_e$: CLARE is the first dedicated machine learning model to predict electron temperature specifically in the plasmasphere (1000–8000 km) using satellite and solar indices.
2. Novel Architecture: It successfully reframes a continuous regression problem as a classification task, demonstrating that discretization improves robustness in noisy, data-scarce environments.
3. Built-in Uncertainty: The model provides confidence metrics without requiring additional computational modules (unlike Bayesian Neural Networks or ensemble methods).
4. Benchmarking: It establishes a new state-of-the-art baseline, significantly outperforming established empirical models (Titheridge and Titheridge-IRI).

4. Results and Performance

The model was evaluated on two held-out test sets: a "quiet" period (randomly sampled) and a "storm" period (Jan 30 – Feb 7, 1991, a known geomagnetic storm).

Performance Metrics (Accuracy within 10% of ground truth):

• Quiet Periods: CLARE achieved 69.67% accuracy, vastly outperforming the Titheridge-IRI baseline (13.49%) and the standard Titheridge model (11.90%).
• Storm Periods: CLARE achieved 46.17% accuracy, compared to 12.26% for Titheridge-IRI. This is significant given that storm data comprised only 0.74% of the training set.
• Statistical Metrics: CLARE achieved an $R^2$ of 0.7829 (quiet) and 0.6647 (storm), with the lowest Root Mean Square Error (RMSE) among all tested models.

Ablation Studies:

• Discrete vs. Continuous: Reverting to a standard continuous regression (MSE loss) reduced accuracy by 6.46% on quiet periods, confirming the superiority of the classification-based approach for this specific task.
• Feature Importance: Solar indices were found to be the dominant predictor during quiet periods. Interestingly, during storm periods, spatial features (spacecraft state) became more predictive than solar indices, suggesting storm dynamics are highly localized or dependent on immediate spacecraft conditions.

5. Significance and Future Directions

Scientific Impact:

• Data-Driven Space Weather: CLARE demonstrates that high-accuracy modeling of the plasmasphere is possible using publicly available data, filling a critical gap in space weather forecasting.
• Handling Data Imbalance: The model shows remarkable resilience in predicting rare events (storms) despite severe data imbalance, though performance drops during storms compared to quiet times.
• Interpretability: The probability distribution output allows operators to assess prediction confidence, a crucial feature for space weather applications where false alarms can be costly.

Limitations and Future Work:

• Storm Performance: While improved, accuracy during storms is lower due to the rarity of such events in the training data. The authors suggest future work could utilize weighted loss functions, data augmentation, or hybrid approaches integrating physics-based priors to further enhance storm prediction.
• Physics Integration: The authors propose merging ML with physics-based models (e.g., using ML to generate synthetic data for physics models or using physics outputs as ML inputs) to push performance boundaries further.

In conclusion, CLARE represents a paradigm shift in plasmaspheric modeling, proving that classification-based regression is a superior strategy for handling the noise and complexity of space physics data compared to traditional continuous regression.