astronomy/ai produced the result/The Astrophysical Journal 2024 · v2
Neural network reads simulated quasar light curves to estimate black hole properties
Researchers trained a latent stochastic differential equation model on simulated ten-year quasar observations. The network filled in gappy, noisy light curves and estimated the physical parameters behind them, replacing a Gaussian process fit used for comparison.
spectrum · one line per step, placed by what the step does · bright lines used AI
Latent Stochastic Differential Equations for Modeling Quasar Variability and Inferring Black Hole Properties
The Astrophysical Journal, 2024
doi:10.3847/1538-4357/ad2988 · record aix-00045 v2 · checked 2026-10-08
- AI was for
- Property prediction, Denoising
- Model family
- Recurrent neural network, Autoencoder, Multilayer perceptron, Gaussian process
- Checked by
- Held-out10000 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
A quasar is the blazing centre of a distant galaxy, where gas spirals into a supermassive black hole and heats up in a flattened disc. Quasars flicker. Their brightness wanders up and down in a way that is not a clean repeating pattern but a random walk, and the flicker arrives at slightly different times in different colours, because the inner, hotter parts of the disc respond before the cooler outer parts. Those time lags carry information about the black hole and its disc. Reading them is hard: telescopes see each object only now and then, through cloud and daylight gaps, with measurement noise on every point.
The researchers set out to recover both things at once from such imperfect records: a smooth, continuous estimate of how a quasar actually varied, and estimates of the physical quantities behind that variation, such as the black hole's mass and the viewing angle of the disc. To test this, they built artificial quasars whose true answers were known. A ray-tracing simulation produced the colour-dependent lags for nine sampled physical parameters, a random-walk signal was fed through them, and the result was degraded to match the observing pattern and noise expected from the Rubin Observatory's ten-year survey.
Where AI came in
The AI was the measuring instrument. A network built from a recurrent encoder, which reads a sequence of daily brightness values with the missing days marked as missing, a decoder based on a stochastic differential equation, which generates a continuous wandering curve, and a small feed-forward estimator for the physical parameters, was trained from scratch on the simulated light curves. It has 903,597 adjustable internal numbers, and training took about six weeks on a single graphics processor.
Given a light curve it had not seen, the trained model returned a reconstruction with an uncertainty at every time step, plus probability distributions for the physical parameters, without the lengthy sampling procedure such fits usually require. On a held-out set of 10,000 light curves, its reconstructions scored better on three error measures than a Gaussian process regression fit, a standard statistical curve-smoothing method, run on the same data. Its parameter estimates were tight for the variability timescale and overall variability amplitude, looser for mass, inclination and disc temperature slope, and weak for redshift, Eddington ratio, spin and corona height.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The authors adapted latent stochastic differential equations to model simulated quasar light curves and to infer black hole and accretion disk properties from them. Training data were simulated by convolving a damped random walk X-ray driving signal with general relativistic ray-traced accretion disk transfer functions, then degraded to LSST-like ten-year cadences and noise. On a held-out test set of 10,000 light curves, the model produced reconstructions with lower RMSE, MAE and negative Gaussian log-likelihood than a fixed-noise multitask Gaussian process regression baseline, and returned calibrated posteriors that constrained the variability timescale and asymptotic structure function well, the black hole mass, inclination angle and temperature slope less tightly, and the redshift, Eddington ratio, spin and corona height only weakly.
How AI was used
A latent SDE model combining a GRU-D-based recurrent encoder, a neural SDE decoder solved with Euler–Maruyama, a recurrent projector and an MLP parameter estimator was trained from scratch on simulated six-band LSST quasar light curves. The simulation chain was non-learned: transfer functions were ray-traced with Sim5 from uniformly sampled physical parameters, convolved with a damped random walk driving signal, and degraded using rubin_sim cadences and LSST noise. The network took brightness and error values at daily intervals with unobserved steps masked, and was trained for 100 epochs with Adam, gradient clipping and KL annealing, using a joint loss combining a negative ELBO, a weighted mean squared error on context points, and the negative log-likelihood of a multivariate Gaussian posterior parameterised by a predicted mean and Cholesky factor over logit-transformed labels. At inference the trained model reconstructed held-out light curves with uncertainties and predicted parameter posteriors directly, without MCMC sampling. A fixed-noise multitask Gaussian process with an absolute exponential kernel and intrinsic co-regionalization, implemented in BoTorch and GPyTorch, was fitted to the same light curves as a baseline.
The shape of the work
Structural · the record, drawn
no AI
Compute accretion disk transfer functions
Numerical or physics simulation, including where a learned surrogate replaces it.
modeled using the general relativistic ray-traced accretion disk simulation software Sim5where the paper describes this · verbatim
no AI
Simulate and degrade mock LSST light curves
Numerical or physics simulation, including where a learned surrogate replaces it.
We impose LSST-like observation cadences and noise by using rubin_sim with the baseline_v2.1_10yrs rolling cadencewhere the paper describes this · verbatim
AI
Train latent SDE on simulated light curves
Fitting model parameters, including fine-tuning an existing model. The AI stood in for statistical model.
We train our network with 100,000 light curves per epoch that are randomly regenerated on the fly each epochwhere the paper describes this · verbatim
AI
Reconstruct light curves and infer parameter posteriors
Running a trained model over new data to predict, classify or score. The AI stood in for statistical model.
we apply it to the test set of light curves that are produced from driving variability and transfer functionswhere the paper describes this · verbatim
AI
Fit multitask GPR baseline to the same light curves
Running a trained model over new data to predict, classify or score.
Our fixed-noise multitask GPR baseline is implemented in BoTorch, based on the GP library GPyTorch.where the paper describes this · verbatim
no AI
Evaluate reconstruction and posterior calibration against truth
Testing outputs against ground truth.
Table 2 compares the light-curve reconstruction performance of our latent SDE model with the GPR baseline across the test setwhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The paper's results are the latent SDE's light-curve reconstructions and parameter posteriors; the reported findings exist only as model outputs compared against a GPR baseline.
we adapt latent SDEs to jointly reconstruct multivariate quasar light curves and infer their physical propertieswhere the paper describes this · verbatim
each metric across our test set of 10,000 light curveswhere the paper describes this · verbatim
The model and code are open source and available atwhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- DataWhether the data are available is not stated.
- Version of Latent SDE (RNN encoder, neural SDE decoder, MLP parameter estimator)Which version of the model was used is not stated.
- Version of Fixed-noise multitask Gaussian process regression baseline (absolute exponential kernel, intrinsic co-regionalization)Which version of the model was used is not stated.
- What step 5 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00045, version 2, checked by a person on 2026-10-08. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error