PaperlayerView exact revision on arXiv

Source-linked AI summary

MAGiDiff: Sampling the Photospheric Vector Field from UV/EUV Filtergrams

Ruoyu Wang, David Fouhey

arXiv:2609.40043v1astro-ph.SRastro-ph.IMcs.CV

TL;DR

Photospheric vector magnetograms rely on demanding full-Stokes observations, while UV/EUV intensity provides an indirect and ambiguous magnetic-field cue. MAGiDiff uses denoising diffusion to estimate vector magnetograms from UV/EUV filtergrams; the authors report plausible structures, promising cross-cycle generalization, and full-disk recovery of main active-region structure, with a need for additional global context.

  • Problem

    Full-Stokes vector magnetograms are demanding to acquire, and estimating photospheric fields from UV/EUV intensity is fundamentally ambiguous.

    Source: §1 · p. 1

  • Method

    MAGiDiff uses denoising diffusion to estimate Hinode/SOT-SP-like heliographic vector magnetograms from SDO/AIA UV/EUV filtergrams.

    Source: §1 · p. 2

  • Results

    MAGiDiff produces visually plausible magnetograms, shows promising cross-cycle generalization, and recovers main active-region structure in full-disk experiments, where additional global context is needed to resolve domain gaps.

    Source: §5 · p. 22

  • Takeaways & Limitations

    The authors frame MAGiDiff as a way to probe information in predictive distributions and obtain supplemental magnetic-field context, not as a replacement for spectropolarimetric measurements.

    Source: §1 · p. 1, §5 · p. 22

    Polarity errors occur in complex active regions and localized plage, while weak-field structures and sharp boundaries can be missed or spatially misaligned.

    Source: §4.2 · pp. 9–10

Abstract

from arXiv · show

Photospheric vector magnetic fields are foundational to modeling, understanding, and forecasting solar activity. These data are usually produced by inverting and disambiguating the full Stokes vector at multiple passbands, which is demanding. Here, we investigate how well we can estimate photospheric vector magnetograms from UV/EUV filtergrams. This problem is challenging and intrinsically ambiguous without polarization information, as the mapping from UV/EUV intensity to the magnetic field is indirect and ill-posed. We introduce MAGiDiff, a machine-learning-based method that uses denoising diffusion models to estimate vector magnetograms from UV/EUV filtergrams. As input, MAGiDiff takes a stack of filtergrams from the Solar Dynamics Observatory (SDO) / Atmospheric Imaging Assembly (AIA); as output, it is trained to estimate the disambiguated vector magnetogram as seen by Hinode / Solar Optical Telescope-Spectro-Polarimeter (SOT-SP). We show that MAGiDiff can accurately mimic the Hinode ground-truth. Additionally, we probe MAGiDiff's understanding of the physical structure and magnetic connectivity. On full-disk, we show that it produces plausible structures for active regions. MAGiDiff generalizes across solar cycles despite hemispheric polarity reversal, and can be fine-tuned to other EUV instruments including STEREO/EUVI and GOES-R/SUVI. While clearly not a substitute for a dedicated instrument, MAGiDiff opens the door to new capabilities.

1. INTRODUCTION

Photospheric vector magnetograms are valuable for solar research but require demanding full-Stokes observations that can limit their cadence, coverage, or availability. MAGiDiff investigates estimating them from more widely available UV/EUV filtergrams despite the indirect and ambiguous relationship between intensity and magnetic field.

1. INTRODUCTION

  • Full-Stokes measurements at multiple wavelengths are demanding, and vector magnetograms can consequently be limited in cadence, coverage, or availability across missions.UV/EUV observations offer a more widely available source for exploring complementary estimates.Source: §1 · p. 1
  • UV/EUV intensity is fundamentally ambiguous as a cue to photospheric magnetic fields, though multiple indirect observations can constrain possible fields.A consistent field flip yields identical intensity images, while coronal filtergrams can trace field lines that connect to the photosphere.Source: §1 · pp. 1–2
  • MAGiDiff uses denoising diffusion models to estimate Hinode/SOT-SP heliographic vector magnetic flux density from SDO/AIA UV/EUV filtergrams.The target components include intrinsic field strength and filling fraction; Hinode/SOT-SP observations are resampled to the SDO/HMI grid.Source: §1 · p. 2
  • The paper evaluates unseen-region and cross-cycle performance, predictive-distribution structure, full-disk generalization, and fine-tuning to other instruments.The authors state that MAGiDiff supplements rather than replaces dedicated spectropolarimetric measurements.Source: §1 · p. 2

2. DATA

The study pairs SDO/AIA UV/EUV filtergrams with Hinode/SOT-SP vector magnetograms, aligns them to the HMI grid, and divides observations by acquisition time. It also prepares separate paired data for fine-tuning to other EUV instruments.

2. DATA

  • Inputs are nine SDO/AIA UV/EUV filtergrams; targets are disambiguated heliographic vector components from Hinode/SOT-SP Stokes inversions.Hinode/SOT-SP scans provide dense sampling, typically over roughly an hour, and yield higher-quality vector magnetograms than SDO/HMI.Source: §2 · p. 2
  • AIA and Hinode/SOT-SP data are co-aligned to the HMI frame, while intrinsic scan-time misalignment is left for the method to handle.Latitude and solar-cycle indicator channels provide positional and temporal context for polarity inference under Hale’s law.Source: §2 · pp. 2–3
  • The dataset contains 71,243 paired observations from 2011–2024; the time-based test set uses 2016 and 2024, with half-year buffers omitted.After crop-size filtering, training, validation, and test sets contain 29,112, 6,137, and 8,307 samples, respectively.Source: §2 · p. 3
  • For STEREO/EUVI and GOES-R/SUVI fine-tuning, EUV filtergrams are paired with reprojected SuperSynthIA magnetograms.SuperSynthIA produces Hinode/SOT-SP-like magnetograms from SDO/HMI data.Source: §2 · p. 3

3. METHOD

MAGiDiff treats vector-magnetogram prediction as conditional sampling from UV/EUV observations and implements it with a latent diffusion model. Its pipeline encodes magnetograms, projects conditioning inputs to latent resolution, trains a denoiser, and samples predicted fields through reverse diffusion.

3. METHOD

  • The task is to sample plausible photospheric vector magnetograms from p(B | I), conditioned on co-registered UV/EUV maps and auxiliary physical context.This formulation supports multiple plausible magnetograms for the same input.Source: §3 · p. 3

3.1. Diffusion Framework

  • MAGiDiff uses latent diffusion: a VAE encodes magnetograms, and a denoiser learns the conditional distribution through forward corruption and reverse denoising.The model predicts a velocity term rather than noise or the clean magnetogram alone; zero terminal SNR and DDIM sampling are also used.Source: §3.1 · pp. 3, 5

3.2. Model Architecture

  • The VAE, spatial rescaler, and denoising U-Net form the model; the VAE compresses magnetograms and the rescaler projects filtergrams to latent resolution.The rescaler combines multi-scale, learnable-convolution, and interpolation branches; the denoising U-Net uses attention and skip connections.Source: Figure 1 · p. 4, §3.2 · pp. 5–6, Figure 2 · p. 5

3.3. Data Normalization

  • A log1p transform and per-channel standardization normalize AIA intensities, while auxiliary channels retain specified encodings and magnetogram components use signed square-root scaling.The magnetic-field scale is 4,000 Mx cm−2; a component at that magnitude has unit magnitude after normalization.Source: §3.3 · p. 6

3.4. Training and Inference Process

  • The conditioning tensor combines nine AIA filters, a disk mask, heliographic latitude, and a solar-cycle indicator with the three heliographic target components.The VAE latent is downsampled by a factor of four in each spatial dimension and has six channels.Source: §3.4 · p. 6

3.4.1. Training of VAE

  • The VAE is trained on 128″ × 128″ Hinode/SOT-SP cutouts using reconstruction and KL-divergence losses, then frozen for denoiser training.The KL weight is set to 10−3 to balance faithful reconstruction and a regularized latent space.Source: §3.4.1 · pp. 6–7

3.4.2. Training of Denoising Network

  • With the VAE frozen, the spatial rescaler and denoising network are trained on paired cutouts to predict velocity using an ℓ2 objective.The rescaled conditioning tensor and noisy magnetogram latent are concatenated before entering the denoiser.Source: §3.4.2 · p. 7

3.4.3. Inference Process

  • Inference starts from Gaussian noise and repeatedly denoises the latent conditioned on UV/EUV inputs before decoding it into a predicted magnetogram.Repeated runs yield distinct samples from the learned conditional distribution, enabling diversity and per-pixel uncertainty assessment.Source: §3.4.3 · p. 7

4. RESULTS

The results assess MAGiDiff’s magnetogram predictions, finding that they reproduce broad field distributions, spatial structure, and relationships while matching or often exceeding a regression baseline in per-pixel accuracy. They also examine predictions across solar cycles, full-disk data, and instruments, noting persistent ambiguity in polarity and limitations in large plage and weakly constrained field strengths.

4.1. Qualitative Results

  • MAGiDiff recovers the locations, spatial distribution, relative strengths, and fine-scale variations of dominant Hinode/SOT-SP structures, while tending to underestimate the strongest fields.The qualitative agreement holds across a broad range of field strengths.Source: §4.1 · p. 9

4.2. Test Set Distributional Agreement

  • Across the test set, MAGiDiff reproduces broad component distributions, but shows opposite-sign and near-zero prediction branches and underestimates extreme fields.The opposite-sign branch reflects polarity errors in complex or localized structures; near-zero predictions arise from missed weak structures or spatial misalignment.Source: §4.2 · pp. 9–10

4.3.1. Effect of the Polarity Prior

  • With the polarity prior, MAGiDiff reaches 92.7% polarity accuracy, versus 86.7% without it; UV/EUV intensities alone are sign-invariant, motivating latitude and solar-cycle inputs.The ±B formulation performs near chance; the prior uses latitude and cycle information to apply Hale’s law.Source: §4.3.1 · p. 10, §4.3 · p. 10, Table 1 · p. 11
  • Classifiers infer hemisphere at 94.7% accuracy and cycle at 98.5%; replacing AIA 304˚A with its dataset mean reduces cycle-classification accuracy to 50%.The authors attribute these implicit cues to geometric effects for hemisphere and residual signal-to-noise changes from instrument degradation for cycle; they favor supplying the prior directly.Source: §4.3.1 · pp. 11–12

4.3.2. Classification by AR type

  • The polarity prior most clearly improves recovery in simple bipolar β regions; gains are smaller in complex βγδ regions, while βδ results are based on only 112 samples from four active regions.Unipolar α regions are also predicted with high accuracy, plausibly because their UV/EUV morphology retains clues to their origin as evolved bipolar remnants.Source: Table 1 · p. 11, §4.3.2 · p. 12

4.4. Comparison with Regression Baseline

  • The regression U-Net gives one deterministic estimate, whereas MAGiDiff samples a conditional distribution; the comparison tests whether stochastic modeling preserves plausible fine-scale structures.The input-output alignment permits image-to-image regression, but indirect UV/EUV cues leave local fields weakly constrained, so an ℓ2 estimate may average plausible configurations.Source: §4.4 · p. 12

4.4.1. Qualitative Comparison

  • Both models recover large-scale fields, but MAGiDiff preserves sharper structures, quiet-Sun texture, and fine-scale power more closely than the smoother regression baseline.The baseline weakens compact features and sharp boundaries, and loses power at finer spatial scales.Source: §4.4.1 · pp. 12, 14, Figure 6 · p. 13, Figure 7 · p. 13

4.4.2. Metrics

  • The comparison metrics cover polarity-aligned pixel accuracy on strong-field pixels and structural fidelity through distributional, gradient, and spectral measures.They include MAE, percentage of errors below 300 Mx cm−2, W1, ∇-ratio, ∇-MAE, and RALSD; strong-field metrics use |αB| > 1000 Mx cm−2.Source: §4.4.2 · p. 14

4.4.3. Quantitative Results

  • In a single-sample comparison, MAGiDiff matches or often exceeds the regression baseline per pixel and more faithfully reproduces Hinode/SOT-SP’s distributions and spatial structure.The comparison uses MAE and the fraction of strong-field pixels below 300 Mx cm−2 for per-pixel accuracy, alongside distributional and structural metrics.Source: Table 2 · p. 16, §4.4.3 · p. 16

4.4.2. Metrics

  • The evaluation measures strong-field per-pixel errors and, over valid pixels, distributional and structural agreement, including gradient-based sharpness and power-spectrum similarity.The radially averaged log-spectral distance is computed on 256×256 crops; predictions are aligned to the ground-truth global polarity before metrics are calculated.Source: §4.4.3 · p. 16, Table 2 · p. 16

4.5. Statistical Characteristics of the Predictive Distribution

  • Correlation maps show intra-plage coherence and relationships largely consistent with magnetic connectivity, while sampled ensembles also provide per-pixel uncertainty estimates.Uncertainty is the per-pixel standard deviation normalized by mean magnitude, computed after downsampling realizations.Source: §4.5 · pp. 16–17

4.6. Cross Solar Cycle Generalization

  • Trained on cycle 24 with polarity-flip augmentation, MAGiDiff recovered correct polarity in 89.7% of 16,201 cycle-25 samples at τ < 1 and 83.3% at τ < 0.5.Test samples span 2020 to 2024.Source: §4.6 · pp. 17–18

4.7. Full Disk Generalization

  • With a lowered-noise schedule and overlapping tiled inference, full-disk predictions recover active-region structure broadly resembling HMI and SuperSynthIA, but large plage can have incorrect polarity without global context.The model may preserve plage location and morphology while assigning part of a bipolar structure the wrong polarity or producing an overly coherent unipolar region.Source: §4.7 · p. 18

4.8. Migration to STEREO/EUVI and GOES/SUVI Data

  • In a proxy evaluation using reprojected SuperSynthIA targets, fine-tuned STEREO/EUVI and GOES/SUVI models recover dominant active-region structures and broadly match field-value distributions, with residual plage errors and strong-field underestimation.The EUV-only inputs provide indirect constraints on the photospheric field, and the proxy setup does not directly validate far-side predictions.Source: §4.8 · pp. 19–20

5. DISCUSSION

MAGiDiff estimates heliographic vector fields from UV/EUV inputs, with cutout predictions that agree with Hinode/SOT-SP and full-disk predictions that recover main active-region structure. The discussion also covers diffusion design choices and limits involving data alignment, full-disk context, and polarity ambiguity.

5. DISCUSSION

  • Full-disk experiments recover main active-region structure, while cutout predictions capture large-scale structure and fine details in strong agreement with Hinode/SOT-SP.The discussion notes that full-disk results also reveal a need for additional global context to resolve domain gaps.Source: §5 · p. 22
  • The breadth of the Hinode/SOT-SP and SDO/AIA archives supports creation of a high-quality training dataset.Hinode/SOT-SP has provided nearly 19 years of high-quality data, while SDO/AIA has supplied high-cadence, well-calibrated full-disk filtergrams since 2010.Source: §5 · p. 22
  • MAGiDiff extends prior UV/EUV-to-magnetogram work by estimating the full heliographic vector field with diffusion and probing learned physical correlations.The paper describes it as the first diffusion-based model to estimate the full heliographic vector field end-to-end from EUV observations, and reports transfer tests with STEREO/EUVI and GOES/SUVI.Source: §5 · p. 22
  • Temporal misregistration, crop-trained full-disk artifacts, and unresolved polarity errors constrain MAGiDiff; full-disk prediction needs additional global context.AIA observations are near-instantaneous, whereas SOT-SP magnetograms are assembled over tens of minutes; polarity failures can persist despite a supplied prior.Source: §5 · pp. 22–23

A.1.1. Forward Process

  • The forward diffusion process adds scheduled Gaussian noise during training until the image approaches pure noise.As the timestep increases, the noisy image retains less information from the clean image.Source: §A.1.1 · p. 26

A.1.2. Reverse Process

  • At inference, iterative denoising reconstructs clean images; MAGiDiff's v-prediction parameterization improves training signal across noise levels.The v-prediction weighting reduces domination of gradients by high- or low-noise extremes.Source: §A.1.2 · p. 27, §A.2 · p. 27

A.3. DDIM sampling

  • DDIM removes per-step noise injection, making the reverse trajectory deterministic and allowing fewer sampling steps to reduce generation cost.It uses the same trained denoising network and can subsample the full chain of roughly 1000 steps, for example to about 50.Source: §A.3 · p. 27

A.4. Latent Diffusion

  • Latent diffusion compresses images with a VAE and performs diffusion in latent space before decoding, reducing computational cost and supporting high-resolution data.The encoder maps images to latents and the decoder reconstructs them; the VAE is trained independently and then frozen.Source: §A.4 · pp. 27–28

B. TRAINING AND INFERENCE PROCESS WALKTHROUGH

The walkthrough organizes MAGiDiff's training and inference into three stages, with a table assigned to each stage.

B. TRAINING AND INFERENCE PROCESS WALKTHROUGH

C. ADDITIONAL EXPERIMENTS DETAILS

The appendix supplies supplementary qualitative and quantitative evaluation material spanning test examples, predictive distributions, and cross-instrument adaptation.

C. ADDITIONAL EXPERIMENTS DETAILS

C.1. Additional Qualitative Results

The appendix adds qualitative cutout examples, including further 2016 test-set predictions and figures documenting selected prediction and ground-truth cases.

C.1. Additional Qualitative Results

  • Four additional 2016 test-set cutouts across varied input shapes are shown as qualitative examples in Figures 13 and 14.The text describes these predictions as accurately estimating Hinode/SOT-SP-like vector magnetograms.Source: §C.1 · p. 29, Figure 13 · p. 30, Figure 14 · p. 31
  • Figure 15 shows representative predictions that contribute to the y = −x branch in Figure 4.Source: Figure 15 · p. 32
  • Figure 16 presents ground-truth αBR for four cutouts shown in Figure 8.The caption lists example dates from January, April, July, and October 2016.Source: Figure 16 · p. 33

C.2. Additional Stochastic Sampling Results

MAGiDiff produces multiple plausible magnetograms from one UV/EUV input, with large-scale structure stable across samples and localized weak-field variation. Best-of-k sampling improves pixel-wise accuracy against Hinode/SOT-SP, while distributional and structural metrics remain largely unchanged.

C.2. Additional Stochastic Sampling Results

  • The evaluation uses one realization per input for comparison with the deterministic regression baseline, although the diffusion model represents a distribution of plausible magnetograms.The section probes that distribution through independent samples and a best-of-k analysis.Source: §C.2 · p. 33, §C.2.2 · p. 33

C.2.1. Qualitative Comparison between Realizations

  • Independent realizations recover nearly identical large-scale structures and polarity across αBR, αBϕ, and αBθ, while varying locally in weak-field regions.The paper interprets these samples as plausible alternatives that vary where signal is weak or uncertain.Source: §C.2.1 · p. 33

C.2.2. Quantitative Comparison between Realizations

  • Best-of-k sampling consistently improves pixel-wise accuracy across all field components, while distributional and structural metrics remain largely unchanged.The analysis selects the realization with the smallest error against Hinode/SOT-SP ground truth; it is a controlled test, not normal inference, where ground truth is unavailable.Source: §C.2.2 · p. 33

C.3. Additional Full-Disk Quantitative Results

The full-disk results compare MAGiDiff predictions with SuperSynthIA references and examine both agreement and stochastic consistency. The plots show strongest agreement in the more active region B, while multiple realizations retain consistent structure with localized variation.

C.3. Additional Full-Disk Quantitative Results

  • Three realizations from the same input retain consistent structure and polarity across all three components while allowing small-scale stochastic variation.Figure 17 marks examples of localized variation with red arrows.Source: Figure 17 · p. 34
  • Best-of-N selection improves per-pixel accuracy with diminishing returns, while a single realization already attains full structural fidelity.Table 6 evaluates N = 1, 4, 16, and 32; its metrics cover strong-field pixel errors and distribution, sharpness, and power-spectrum similarity.Source: Table 6 · p. 34
  • Most full-disk and regional prediction density lies near the 1:1 line, with strongest agreement in active region B.Figure 18 reports Pearson correlation and fit slope; pixels with ground-truth absolute value below 150 Mx cm−2 are omitted.Source: §C.3 · p. 33, Figure 18 · p. 35

C.4. Cross-Instrument Fine-Tuning Pipeline

MAGiDiff is adapted to other EUV instruments through preprocessing and fine-tuning while keeping its learning problem fixed. On STEREO/EUVI cutouts, the fine-tuned model produces reasonable estimates, and best-of-k sampling improves performance.

C.4. Cross-Instrument Fine-Tuning Pipeline

  • Preprocessing reprojects and resamples each instrument’s filtergrams, calibrates intensities toward the SDO/AIA range, and sets unavailable AIA channels to zero.The model receives SDO/AIA-like EUV inputs along with an on-disk mask, heliographic latitude map, and solar-cycle indicator.Source: §C.4 · p. 35
  • Because Hinode/SOT-SP does not densely co-observe STEREO/EUVI or GOES/SUVI, fine-tuning uses SuperSynthIA magnetograms as targets and flips the αBθ sign to match Hinode’s convention.The sign flip follows reprojection because SuperSynthIA uses the SDO/HMI αBθ convention.Source: §C.4 · pp. 35–36
  • On the more challenging STEREO/EUVI cutouts, fine-tuned MAGiDiff produces reasonable estimates, and best-of-k sampling improves performance.The evaluation uses SuperSynthIA as the reference.Source: Table 7 · p. 36
  • Fine-tuning first adapts the VAE to SuperSynthIA magnetograms, then trains the denoising U-Net on instrument-specific pairs with the VAE frozen.A sign-flip augmentation also reverses target magnetogram polarity and solar-cycle parity with probability 0.5, leaving EUV filtergrams and latitude unchanged.Source: §C.4 · p. 36

C.4.1. Adaption to STEREO/EUVI

For STEREO/EUVI, MAGiDiff is fine-tuned using paired EUV inputs and SuperSynthIA targets, with time-varying intensity calibration to address photometric mismatch and degradation. The resulting model gives reasonable estimates on held-out EUVI cutouts, but is expected to be most reliable near the 2022–2024 fine-tuning period.

C.4.1. Adaption to STEREO/EUVI

  • The EUVI training set uses 171, 195, and 304 Å channels and contains 4,949 paired observations, split into training, validation, and test periods.The observations are restricted to 2022–2024, when STEREO-A/EUVI substantially overlaps SDO/HMI, enabling SuperSynthIA target construction.Source: §C.4.1 · p. 36
  • Time-varying quantile mapping in log(1 + x) space addresses EUVI photometric mismatch and mission-long sensitivity changes before standard AIA normalization.Matched percentile anchors are aggregated into smoothed, monotonic lookup curves with 101 breakpoints, then interpolated for each intensity value.Source: §C.4.1 · pp. 36–37
  • The correction does not ensure equal reliability across observation times; the model is expected to be most reliable near 2022–2024, and earlier use may require additional adjustment.The stated boundary reflects the fine-tuning dataset’s restricted time range and long-term EUVI intensity changes.Source: §C.4.1 · p. 37
  • On challenging STEREO/EUVI cutouts, fine-tuned MAGiDiff produces reasonable estimates, and best-of-k sampling improves performance.Performance is quantified on a held-out test set using SuperSynthIA as the reference.Source: §C.4.1 · p. 37

C.4.2. Adaption to GOES/SUVI

MAGiDiff is fine-tuned for GOES–16/SUVI and GOES–18/SUVI using overlapping AIA passbands, then evaluated against SuperSynthIA on held-out data. The section describes the paired-data preparation, temporal splits, and satellite-specific normalization.

C.4.2. Adaption to GOES/SUVI

  • The dataset uses the five GOES/SUVI passbands overlapping AIA, requires observations within 10 minutes, and samples Level-2 products at four-hour intervals.Channels are coaligned to the 171 Å frame, upsampled 4.17× to AIA’s angular pixel scale, and paired with reprojected SuperSynthIA magnetograms.Source: §C.4.2 · p. 37
  • The GOES–16/SUVI and GOES–18/SUVI fine-tuning dataset spans 2022–2024 and contains 7,475 pairs, split into 3,160 training, 1,231 validation, and 2,208 test samples.Training uses January 2022–June 2023, validation July–December 2023, and testing April–December 2024.Source: §C.4.2 · p. 37
  • GOES–16/SUVI and GOES–18/SUVI inputs are normalized separately by channel because the satellites are not fully intercalibrated, then combined for fine-tuning.Inputs use log(1+x) normalization, with each sample normalized using parameters for its source satellite.Source: §C.4.2 · p. 37
  • The held-out test set is evaluated against SuperSynthIA, with Figure 19 comparing full-disk prediction and reference pixel-value histograms across instruments and vector-field components.The caption orders columns as αBR, αBϕ, αBθ, and |αB|, and rows as STEREO/EUVI, GOES–16/SUVI, and GOES–18/SUVI.Source: §C.4.2 · p. 38, Figure 19 · p. 38

C.4.3. Additional Quantitative Results

Figure 19 adds a quantitative distributional analysis of the full-disk predictions shown in Figure 11. It compares MAGiDiff prediction distributions with the corresponding SuperSynthIA reference distributions.

C.4.3. Additional Quantitative Results

  • Figure 19 adds a quantitative analysis of the full-disk predictions shown in Figure 11.It compares the MAGiDiff prediction distributions with corresponding SuperSynthIA magnetograms.Source: §C.4.3 · p. 38
  • The comparison is distributional, using full-disk predictions and corresponding reference magnetograms.Source: §C.4.3 · p. 38
  • The figure assesses prediction distributions against the corresponding SuperSynthIA reference distributions.Source: §C.4.3 · p. 38

Related papers

Loading 2609.40043v1…