Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Measuring the Vertical Structure of Active Galactic Nuclei Disks with Transformer Models and the Vera C. Rubin Observatory

Amy Secunda, Sebastian Wagner-Carena, Helen Qu, Shirley Ho

TL;DR

This work tackles mapping the vertical structure of AGN accretion disks by detecting long negative lags in Rubin Observatory light curves. It introduces a transformer-based method trained on a large suite of realistic mocks to infer a three-component posterior for short and long lags, without being misled by missing data. On mocks, the approach achieves 96% recall with 0.04% contamination and 98% accuracy for long-lag predictions, outperforming traditional methods ICCF (54% accuracy) and javelin (21%), while being orders of magnitude faster. This enables scalable exploitation of Rubin data to probe disk inflow timescales and height, providing a powerful test of thin-disk versus magnetically elevated disk models; however, it remains validated on simulations and will need to incorporate correlations between parameters in future work.

Abstract

Reverberation mapping is one of the main techniques used to study active galactic nuclei (AGN) accretion disks. Traditional continuum reverberation mapping uses short lags between variability in different wavelength AGN light curves on the light crossing timescale of the disk to measure the radial structure of the disk. The harder-to-detect long negative lag measures lags on the longer inflow timescale, opening up a new window to mapping out the vertical structure of AGN disks. The Vera Rubin Observatory, with its 6 wavebands, long baseline, and high cadence, will revolutionize our ability to detect short and long lags. However, many challenges remain to detect these long lags, such as seasonal gaps in Rubin light curves, the weak signal strength of the long lag relative to the short lag, and the enormous influx of data for millions of AGN from Rubin. Machine learning techniques have the potential to solve many of these issues, but have yet to be applied to the long negative lag problem. We develop and train a transformer-based machine learning model to detect long and short lags in mock Rubin AGN light curves. Our model identifies whether a light curve in our test set has a long negative lag with 96% recall and 0.04% contamination, and is 98% accurate at predicting the true long lag. This accuracy is an enormous improvement over two baseline methods we test on the same mock light curves, the interpolated cross correlation function and javelin, which are only 54% and 21% accurate, respectively.

Measuring the Vertical Structure of Active Galactic Nuclei Disks with Transformer Models and the Vera C. Rubin Observatory

TL;DR

This work tackles mapping the vertical structure of AGN accretion disks by detecting long negative lags in Rubin Observatory light curves. It introduces a transformer-based method trained on a large suite of realistic mocks to infer a three-component posterior for short and long lags, without being misled by missing data. On mocks, the approach achieves 96% recall with 0.04% contamination and 98% accuracy for long-lag predictions, outperforming traditional methods ICCF (54% accuracy) and javelin (21%), while being orders of magnitude faster. This enables scalable exploitation of Rubin data to probe disk inflow timescales and height, providing a powerful test of thin-disk versus magnetically elevated disk models; however, it remains validated on simulations and will need to incorporate correlations between parameters in future work.

Abstract

Reverberation mapping is one of the main techniques used to study active galactic nuclei (AGN) accretion disks. Traditional continuum reverberation mapping uses short lags between variability in different wavelength AGN light curves on the light crossing timescale of the disk to measure the radial structure of the disk. The harder-to-detect long negative lag measures lags on the longer inflow timescale, opening up a new window to mapping out the vertical structure of AGN disks. The Vera Rubin Observatory, with its 6 wavebands, long baseline, and high cadence, will revolutionize our ability to detect short and long lags. However, many challenges remain to detect these long lags, such as seasonal gaps in Rubin light curves, the weak signal strength of the long lag relative to the short lag, and the enormous influx of data for millions of AGN from Rubin. Machine learning techniques have the potential to solve many of these issues, but have yet to be applied to the long negative lag problem. We develop and train a transformer-based machine learning model to detect long and short lags in mock Rubin AGN light curves. Our model identifies whether a light curve in our test set has a long negative lag with 96% recall and 0.04% contamination, and is 98% accurate at predicting the true long lag. This accuracy is an enormous improvement over two baseline methods we test on the same mock light curves, the interpolated cross correlation function and javelin, which are only 54% and 21% accurate, respectively.
Paper Structure (9 sections, 4 equations, 4 figures)

This paper contains 9 sections, 4 equations, 4 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Simulating Mock Light Curves
  4. Transformer Model
  5. Other Lag Detection Methods
  6. Results
  7. Transformer Model Performance
  8. Comparison to Other Methods
  9. Discussion and Summary

Figures (4)

  • Figure 1: The distribution of the predicted likelihood of a long negative lag ($1-p_\phi$) in our test set mock light curves without (in brown) and with (in blue) long negative lags. The dashed black line shows our cut off, $(1-p_\phi)>0.906$, for filtering out light curves without long lags. This cut-off gives us a sample with $0.04\%$ contamination and 96% completeness.
  • Figure 2: The top panel shows the duration of the true long lag versus the duration of the long lag predicted by our transformer model for our test set of $2.0e4$ mock light curves. The middle and bottom panel show the duration of the true long lag versus the duration of the long lag predicted by the ICCF and javelin, respectively, for a subset of 2000 light curves from our test set. The solid lines are the one-to-one lines for a perfect prediction, the dashed lines show $\pm 100$ days. In the top panel the color bar shows the variance of the posterior for the logarithm of the long lag. In the middle and bottom panels the color bar shows the logarithm of the ratio of the amplitude of the long lag to the amplitude of the short lag. The horizontal gray lines in the bottom panel show odd integer multiples of 180 days which is roughly the seasonal gap in observations. 98% of long lags predicted by our model are within 20% of the true long lag. On the other hand, javelin experiences severe aliasing problems and does not reliably predict the long lag, while the ICCF fails to predict any long lag for light curves with low amplitude ratios.
  • Figure 3: The fraction of light curves with long negative lags detected in each amplitude ratio bin for our transformer model (in blue) and the ICCF method (in purple). The transformer model detects long lags for all amplitude ratios only missing 19% of long lags for the lowest amplitude ratio bin, while the ICCF fails to detect 87% of long lags for light curves with amplitude ratios, $r<0.01$.
  • Figure 4: Our TARP calculation (in blue) of the expected coverage probability versus the credibility level for our model posterior. The dashed black line indicates perfect calibration. Our model does very well on this calibration test, with slight under-confidence in a few quantiles.