🔭 astrophysics

BT2PB: High-cadence polarized brightness of the solar corona from LASCO C2 total brightness observations using deep learning

This paper introduces BT2PB, a deep learning model that utilizes a convolutional neural network trained on decades of LASCO C2 data to predict high-cadence polarized brightness (pB) from total brightness (Bt) observations, thereby enabling new high time-resolution studies of solar coronal dynamics and improving mass estimates for events like coronal mass ejections.

Original authors: Huw Morgan, Harshita Gandhi

Published 2026-10-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Huw Morgan, Harshita Gandhi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Sun is not a static ball of fire; it is a dynamic, churning star surrounded by an atmosphere that stretches millions of miles into space. This outer layer, known as the corona, is a tenuous gas of superheated particles that constantly streams away from the Sun, carrying magnetic fields and energy that can disrupt satellites and power grids on Earth. To understand this atmosphere and predict when it might erupt, scientists need to measure its density and structure with great precision. However, looking at the corona is like trying to see a faint candle flame next to a blinding spotlight. The light we see from the corona is a mix of two things: the actual glowing gas we want to study, and a bright, hazy background caused by sunlight scattering off dust and stray light bouncing off the telescope itself.

For decades, the Large Angle and Spectrometric Coronagraph (LASCO) instrument aboard the SOHO spacecraft has been watching this solar atmosphere. It takes two types of pictures. The first type, called total brightness, is taken very frequently—dozens of times a day. These images are bright and clear, showing the shape of solar eruptions as they happen, but they are contaminated by that bright background haze, making it hard to measure the true density of the gas. The second type, called polarized brightness, is much cleaner. By using special filters that block the hazy background light, these images reveal the true structure of the corona. The problem is that these clean images are taken very rarely, perhaps only once or twice a day. This slow pace means scientists often miss the rapid changes in the corona or cannot build a complete 3D map of it, because they lack enough snapshots to fill in the gaps.

A team of researchers at Aberystwyth University has developed a new way to bridge this gap. They created a computer program based on deep learning, a type of artificial intelligence that learns by studying vast amounts of data. The program, named BT2PB, was trained on thousands of pairs of images taken over the last twenty-five years. For every pair, the program studied a "dirty" total brightness image and its corresponding "clean" polarized brightness image taken at nearly the same time. By analyzing these examples, the computer learned to recognize the subtle patterns that distinguish the true coronal gas from the background noise. It learned to look at a frequent, noisy image and predict what the clean, high-quality version would look like if it had been taken at that exact moment.

When the researchers tested this new tool, they found it worked remarkably well. The program successfully converted the frequent, noisy images into clean, high-quality predictions of the corona's true brightness. On average, the predictions were accurate to within about eight to twelve percent of the actual measurements. The researchers noted that some of this small error came not from the computer's mistake, but from the inherent noise in the real, physical images used for testing, which are often grainy and imperfect. Even with this limitation, the improvement is significant. The tool allows scientists to generate a high-quality view of the corona as often as the telescope takes a picture—tens of times a day—rather than waiting for the rare, clean snapshots.

This capability opens up new possibilities for studying the Sun. Because the program can provide a steady stream of clean data, scientists can now track fast-moving events, such as massive eruptions of gas and magnetic fields, with a level of detail that was previously impossible. They can also use these frequent, high-quality images to build three-dimensional maps of the solar atmosphere, a technique called tomography, which requires many observations from different angles over time. Previously, the slow pace of the clean images meant these maps were often incomplete or impossible to make for many periods. Now, with the ability to generate clean data on demand, researchers can study the Sun's structure and behavior with a clarity and speed that brings us closer to understanding how our star influences the space around us.

Technical Summary: BT2PB – High-cadence polarized brightness of the solar corona from LASCO C2 total brightness observations using deep learning

Problem Statement
The Large Angle and Spectrometric Coronagraph (LASCO) C2 instrument aboard the Solar and Heliospheric Observatory (SOHO) has provided continuous observations of the solar corona for over three decades. While LASCO C2 collects tens of total brightness ( $B_t$ ) observations daily, these images contain a mixture of the polarized K-corona ( $B_k$ ), the unpolarized F-corona, background stars, and instrumental stray light. Consequently, $B_t$ data are unsuitable for quantitative coronal science, such as density inversion or mass estimation, without rigorous calibration and background subtraction.

Conversely, LASCO C2 produces polarized brightness ( $pB$ ) observations, which isolate the K-corona by suppressing F-corona and stray light components. However, $pB$ observations are typically acquired only once or twice per day. This low cadence prevents high-resolution temporal studies of dynamic events (e.g., Coronal Mass Ejections) and limits the efficacy of coronal tomography, which requires dense temporal sampling to reconstruct 3D density structures. Existing methods to derive $B_k$ from $B_t$ often rely on complex, computationally expensive background modeling over long time windows, which becomes cumbersome during spacecraft roll maneuvers.

Methodology
The authors present BT2PB, a Convolutional Neural Network (CNN) framework designed to estimate $pB$ directly from $B_t$ observations, effectively increasing the cadence of calibrated polarized data to match the high-frequency $B_t$ acquisition rate.

Data Preparation: The model was trained on 24,161 paired observations of LASCO C2 $B_t$ and $pB$ sequences collected between January 2000 and March 2025. Data were filtered to ensure temporal proximity (within 30 minutes) and specific instrument configurations (orange filter).
Input Representation: Rather than processing full 2D images, the network operates on 1D brightness profiles extracted at 30 heliocentric distances (logarithmically spaced between $2.3$ and $6.0 R_\odot$ ) and 360 position angles. These profiles are resampled to 128 position angle bins.
Target Quantity: The network predicts the logarithmic ratio $R = \log_{10}(pB/B_t)$ rather than $pB$ directly. This formulation stabilizes training across the wide dynamic range of brightness values. The final $pB$ is reconstructed via $pB = 10^{\log_{10}(B_t) + R}$ .
Network Architecture: The model is a 1D CNN comprising:
- An Encoder with two 1D convolutional layers.
- Four Residual Blocks (ResBlock1D) for feature refinement.
- Metadata Injection: An 8-element metadata vector (including spacecraft roll angle, heliocentric distance, heliographic latitude, and exposure time) is projected via Multi-Layer Perceptrons (MLPs) and added to the feature tensor. This allows the network to account for time-varying instrumental and orbital effects.
- A Decoder projecting features back to the output shape.
- A Physics-Informed Penalty: A loss term penalizes unphysical predictions where $pB > B_t$ (i.e., $R > 0$ ).
Training: The network was trained on 17,099 pairs using the AdamW optimizer with a composite loss function (MSE on $R$ plus the physics penalty). Training was performed on a standard desktop CPU, converging in approximately 73 epochs.

Key Results

Accuracy: On a held-out test set of 2,279 pairs spanning two solar cycles, the model achieves a Mean Absolute Error (MAE) of approximately 8–12% in predicting $pB$ across heliocentric heights from $2.3$ to $6.0 R_\odot$ .
Error Decomposition: The authors estimate that roughly 4% of the total MAE arises from noise and artifacts inherent in the observed $pB$ data (used as ground truth), rather than model failure. Thus, the error attributable solely to the BT2PB prediction is estimated at 4–8%.
Solar Cycle Dependence: Performance varies slightly with the solar cycle, with lower errors generally observed during solar minimum, particularly in equatorial streamer regions.
Robustness to Roll: The model effectively handles the 180° spacecraft roll maneuvers. Ablation studies indicate that the network implicitly learns the roll state from the asymmetry of stray light patterns in the $B_t$ profiles, rendering the explicit roll angle metadata contribution negligible (+0.18% MAE increase when removed).
Case Studies: Visual comparisons for both solar maximum (2012) and solar minimum (2008) conditions show excellent agreement in coronal structure, with the predicted $pB$ capturing major features and streamers, though the model produces smoother results than the noisy observed $pB$ .

Significance and Applications
The paper claims that BT2PB enables new scientific capabilities for the existing LASCO C2 dataset without requiring new hardware or complex background modeling:

High-Cadence Dynamic Studies: By converting the frequent $B_t$ observations into calibrated $pB$ , the method allows for high time-resolution analysis of dynamic events like CMEs, which are often missed by the sparse native $pB$ cadence.
Improved Mass and Density Estimates: The higher temporal resolution facilitates more reliable mass estimates for dynamic events and improves coronal density diagnostics.
Coronal Tomography: The authors demonstrate an application where BT2PB-generated $pB$ data (320 images over 14 days) is used as input for a tomography module (CorTom-ML). This provides full-coverage density maps that would be impossible with the native daily $pB$ cadence (14 images).
Operational Efficiency: The method is computationally efficient, requiring only minutes to train on a standard CPU and offering a streamlined alternative to long-term background subtraction techniques.

The authors conclude that BT2PB provides a highly efficient pathway to unlock high-temporal-resolution, calibrated polarized data from decades of LASCO C2 $B_t$ observations, supporting both scientific research and operational space weather forecasting.

More like this