🔭 astrophysics

Probabilistic Interpolation of Sagittarius A*'s Multi-Wavelength Light Curves Using Diffusion Models

This paper introduces the first application of score-based diffusion models and transformer architectures to probabilistically interpolate sparse, multi-wavelength light curves of Sagittarius A*, demonstrating superior accuracy and calibrated uncertainty estimates compared to traditional Gaussian Processes for reconstructing the black hole's variability.

Original authors: Gabriel Sasseville, Julie Hlavacek-Larrondo, Daryl Haggard, Alexandre Adam, Hadrien Paugnat, Gunther Witzel

Published 2026-01-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gabriel Sasseville, Julie Hlavacek-Larrondo, Daryl Haggard, Alexandre Adam, Hadrien Paugnat, Gunther Witzel

Original paper licensed under CC BY 4.0 (http://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 Big Picture: Filling in the Blanks of a Cosmic Puzzle

Imagine you are trying to listen to a song, but the radio signal is terrible. The music cuts in and out, sometimes the bass is loud, sometimes the treble is missing, and there are long stretches of static. You know a song is playing, but you can't hear the full melody.

This is exactly the problem astronomers face with Sagittarius A* (Sgr A*), the supermassive black hole at the center of our galaxy. Scientists observe it using four different "ears" (telescopes) tuned to different parts of the light spectrum: X-rays, near-infrared, infrared, and sub-millimeter waves.

However, these telescopes don't work perfectly together:

They look at the black hole at different times (irregular scheduling).
They have gaps in their data (like the radio cutting out).
The "noise" (static) varies wildly between the different telescopes.

The goal of this paper is to build a smart computer program that can listen to these broken, messy recordings and reconstruct the full, continuous song of the black hole's activity, while also telling us how confident it is about the parts it had to guess.

The Problem: Why Old Methods Struggle

The authors tried three different ways to solve this "fill-in-the-blanks" puzzle:

The "Smooth Painter" (Gaussian Process): Imagine an artist who loves smooth curves. If they see two dots, they draw a gentle, straight line between them. This method is safe and reliable, but it's too smooth. It misses the sharp, sudden "flares" (bright bursts of energy) that the black hole actually produces. It's like trying to draw a jagged lightning bolt with a soft, round brush.
The "Pattern Matcher" (Transformer): This is a modern AI that is great at spotting patterns in sequences (like how Netflix recommends movies). It's fast and good at connecting dots, but it sometimes gets overconfident. It might guess a pattern exists even when the data is too sparse to support it, leading to "hallucinations" (making up flares that didn't happen).
The "Denoising Sculptor" (Diffusion Model): This is the new method the authors introduced. Imagine a sculpture covered in thick fog. You know the shape is underneath, but you can't see it clearly. The diffusion model works by starting with pure static (fog) and slowly, step-by-step, removing the noise to reveal the shape underneath. It learns what the "real" black hole signal looks like by studying thousands of simulated examples first.

The Solution: The "Denoising Sculptor" (Diffusion Model)

The authors created a new tool called a Diffusion Model specifically for time-series data. Here is how it works in simple terms:

Training: They didn't just feed it real data (which is too messy). Instead, they fed it 16,000 simulated movies of the black hole. These simulations were built to look exactly like the real thing, including the gaps and the noise.
The Process: The model learns to take a messy, incomplete signal and "clean it up." It doesn't just draw a smooth line; it learns the rhythm and chaos of the black hole.
The Magic: When it reconstructs the data, it doesn't just give one answer. It gives a range of possibilities with a confidence score. It knows when it's guessing and when it's sure.

The Results: Who Won the Contest?

The authors tested all three methods on both the fake (simulated) data and the real telescope data.

Accuracy: The Diffusion Model was the clear winner. It was the best at capturing the sharp, sudden flares of the black hole without smoothing them out. It made the fewest mistakes in guessing the missing numbers.
The "Hallucination" Test: In one specific test, the "Pattern Matcher" (Transformer) and the "Smooth Painter" (Gaussian Process) both invented a fake flare in the X-ray data around 11:30 AM. The real data showed nothing happened. The Diffusion Model correctly stayed quiet, showing it understood that a flare in one color of light doesn't always mean a flare in another.
Confidence: The Diffusion Model was very good at saying, "I'm not sure about this part," especially in the noisier sub-millimeter data. The other models were either too unsure (too wide of a guess) or too confident (too narrow of a guess).

The Catch: What the Model Can't Do Yet

The paper is honest about its limits:

The "Simulator" Bias: The model was trained entirely on computer simulations. While these simulations are very good, they are still just one version of reality. If the real black hole behaves in a way the simulation didn't predict, the model might struggle.
The Noisiest Band: Even the best model struggled the most with the sub-millimeter data because that data is so sparse and noisy. It's like trying to hear a whisper in a hurricane; even the best AI has a hard time there.

The Bottom Line

This paper introduces a new, powerful way to reconstruct the "story" of our galaxy's black hole from broken, messy data. By using a diffusion model (a type of AI that learns by removing noise), the authors can fill in the gaps in telescope observations more accurately than before.

Most importantly, this new tool doesn't just guess; it tells us how much we should trust its guess. This helps astronomers avoid making mistakes by believing in "ghost" flares that the computer made up, leading to a clearer understanding of how black holes actually behave.

Technical Summary: Probabilistic Interpolation of Sagittarius A's Multi-Wavelength Light Curves Using Diffusion Models*

Problem Statement
Understanding the variability of Sagittarius A* (Sgr A*) requires coordinated, multi-wavelength observations spanning the electromagnetic spectrum (X-ray, near-infrared, infrared, and submillimeter). However, these observations suffer from irregular sampling, band-dependent noise, and substantial data gaps due to telescope scheduling and logistical constraints. These limitations hinder the robust identification of flares and the measurement of cross-band time lags, which are critical diagnostics for the physical processes driving accretion and emission near the black hole's event horizon. The core challenge is an inverse modeling problem: reconstructing a continuous, underlying signal $x(t)$ from sparse, noisy, and irregularly sampled observations $\{y(t_k)\}$ while quantifying the uncertainty inherent in unobserved regions.

Methodology
The authors propose a diffusion-based generative framework for interpolating multivariate astrophysical time series. The study compares three distinct modeling approaches:

Multi-Output Gaussian Process (MOGP): Used as a principled probabilistic baseline. The authors implement an Intrinsic Coregionalization Model (ICM) using a Radial Basis Function (RBF) kernel for temporal covariance and a low-rank structure for cross-band correlations. This approach assumes Gaussianity and smoothness but scales poorly with data dimensionality.
TripletFormer: A transformer-based architecture (Yalavarthi et al. 2023) capable of predicting uncertainties. It utilizes an encoder-decoder structure with induced multi-head attention and cross-attention blocks to learn inter- and intra-channel correlations. It outputs parameters for a Gaussian distribution to represent predictive uncertainty.
Continuous Stochastic Process Diffusion (CSPD): The primary contribution of this work. This model extends score-based diffusion models to continuous stochastic processes.
- Forward Process: Data is progressively corrupted by noise drawn from a Gaussian Process (GP) with a temporal kernel, rather than independent Gaussian vectors. This preserves temporal structure during corruption.
- Reverse Process: A neural score network learns to reverse the diffusion, conditioned on both the noisy state and the observed values.
- Interpolation Mechanism: During inference, observed values are clamped (fixed) at each diffusion step, while missing entries are progressively denoised. This allows the model to handle irregular sampling without predefined analytic bases (e.g., Fourier or wavelets).

Data and Training

Observational Data: Real multi-wavelength data from a July 2019 campaign involving Chandra (X-ray), GRAVITY (NIR), Spitzer (IR), and ALMA (submm). This dataset serves as a case study for qualitative evaluation.
Simulated Data: The models are trained and quantitatively evaluated on 16,350 simulated 24-hour light curves generated using a semi-empirical radiative model (Witzel et al. 2021). These simulations capture correlated multi-band variability, realistic observation cadences, and wavelength-specific noise. A masking procedure replicates the irregular sampling of real campaigns.

Key Contributions

First Application of Diffusion Models: This work presents the first application of score-based diffusion models to astronomical time series interpolation.
Calibrated Uncertainty: It introduces the first transformer-based model for light curve reconstruction that includes calibrated uncertainty estimates, addressing a gap in existing transformer applications for this domain.
Probabilistic Framework: The CSPD model provides a flexible, data-driven approach to interpolation that captures complex temporal and inter-band dependencies without assuming a fixed parametric prior.

Results

Performance on Simulated Data: The CSPD model achieved superior accuracy (lowest Mean Squared Error and Continuous Ranked Probability Score) across all wavelength bands compared to MOGP and TripletFormer. It demonstrated particular strength in capturing sharp, bursty variability in X-ray and NIR bands, which the smoother MOGP and TripletFormer tended to over-smooth.
Robustness to Missing Data: In stress tests involving random and burst (contiguous) missing data, CSPD maintained the best accuracy across most sparsity levels. While TripletFormer showed competitive performance in random missingness, it suffered significant degradation under burst missingness. MOGP remained stable but conservative.
Performance on Real Data: When applied to real Sgr A* observations (without fine-tuning), CSPD and TripletFormer tracked rapid flares more effectively than MOGP. Notably, CSPD avoided "hallucinating" false-positive X-ray flares that appeared in the other models' predictions during specific intervals, suggesting it better internalized the physical decoupling between NIR and X-ray emission mechanisms.
Uncertainty Calibration: MOGP provided the most conservative and well-calibrated uncertainty estimates. CSPD showed slightly narrower credible intervals (mild underestimation of uncertainty) but remained well-calibrated. TripletFormer exhibited significant undercoverage (overconfidence), particularly in the highly variable NIR band.

Significance and Claims
The authors claim that diffusion models offer a promising avenue for high-fidelity, uncertainty-aware reconstruction of multi-wavelength variability in Sgr A*. The CSPD model successfully generalizes from simulated training data to real observational data, demonstrating the ability to reconstruct missing data while respecting the complex, non-Gaussian nature of astrophysical variability.

The paper modestly concludes that while the results are encouraging, the problem remains intrinsically difficult, particularly in low signal-to-noise bands like submillimeter. The authors emphasize that their work advances the ability to study multi-scale variability of accreting black holes even in the absence of complete observations. They note that future work must address generalization to diverse physical scenarios (beyond the single simulator used) and incorporate more sophisticated domain adaptation techniques to bridge the gap between synthetic and real data distributions. The framework is presented as a tool to extract more information from incomplete datasets, potentially aiding in the measurement of spectral indices, time lags, and power spectra.