BlaST

BlaST (Blazar Synchrotron Tool) is a tool to estimate the synchrotron peak of blazars given its spectral energy distribution. The package is available on pip and can simply be installed via

pip install astroblast

Running the tool is as simple as typing

blast sed.txt

where sed.txt is a output file generated by the VOUBlazar tool and might look something like the following:

   1  matched source  227.16230  -49.88400  99
 Frequency     nufnu     nufnu unc.  nufnu unc. start time   end time   Catalog     Reference
    Hz       erg/cm2/s     upper       lower        MJD         MJD   
---------------------------------------------------------------------------------------------------------------------------
 2.418E+17   2.185E-13   3.139E-13   1.230E-13  55000.0000  55000.0000  RASS        Boller et al. 2016, A&A, 103, 1                                                                                                                                                                         
 2.418E+17   5.085E-13   6.281E-13   3.889E-13  58150.0000  58150.1016  OUSXB       Giommi et al. 2019, Accepted for publication in A&A  

This gives the following output:

12.42 (+/- 0.71)

The first value is the synchrotron peak and the second the 95% prediction interval both in log10. It is possible to specify a different prediction interval width by the -w argument, followed by the width in sigmas:

blast sed.txt -w 1.0

BlaST also supports bulk estimation, e.g. estimating a whole catalogue, by providing a directory or zip file containing the seds as well as an output file in which the results will be written to as csv:

blast seds.zip -o estimates.csv

With V2 you now also have the option to estimate the synchrotron peak flux (erg/cm2/s), again in log10. You enable this with the -f flag:

-12.39 (+/- 0.63) @ 12.39 (+/- 0.39)

Changes from v1 to v2

With v2 there has been some major changes. Some are listed in the following:

• New data set (courtesy of N. Krieger)
  The data set has been refined resulting in a new set independent of the previous one. This one now also includes labels for the flux.
• Flux estimate
  Biggest difference is the estimation of the synchrotron peak's flux by an independent ensemble of neural networks.
• Save models in ONNX format
  Previously, the tool needed pytorch installed, which is quite large. We moved to ONNX runtime to make the installation slimmer.

How it works

BlaST consists of an ensemble of similar neural networks power by pytorch based on the method presented in Lakshminarayanan et. al. 2016. The data set (see SEDs.zip) is based on the three catalogues 4LAC, 3HSP and 5BZCat and divided into 5 bags. For each bag an ensemble of 5 models are trained totalling 25 independent models.

The bagging allows the reapply the ensemble on the training data as an out-of-bag estimation, i.e. excluding the ensemble members that were trained on a training set without that specific bag. The results are shown in estimates.csv.

For specific details, e.g. how the data was enhanced and how the bias in the data set was handles, take a look in BlaST.ipynb which is the notebook used to train the models. Note that if you want to run that notebook you'll also need to install PyTorch Lightning

Performance

The performance was evaluated using out-of-bag estimates on the training data including a 95% prediction interval. Metrics are shown as median and 25%/75% quantile. Since the model for predicting the peak frequency and peak flux are independent, metrics are shown for both.

Metric	Value	Freq	Flux
Abs Err	Median	0.163	0.104
	25%	0.075	0.049
	75%	0.314	0.186
PI Width	Median	1.267	1.087
	25%	1.009	0.971
	75%	1.625	1.198
Interval	Median	1.400	1.088
Score	25%	1.010	0.972
	75%	1.636	1.199

Peak Frequency Estimate

The following show the prediction histogram with the median as black line, as well as the 90% and 10% quantile shown as dotted lines

The next shows the prediction interval (95%) widths. The percentages above and below show the respective amount of samples outside their intervals.

The final one shows the prediction error and the prediction interval distribution using a moving window and smoothed using cubic splines.

Peak Flux Estimate

Same plots, but now for the peak flux estimate.

Use BlaST in code

BlaST can also be imported as python package:

from blast import *

sed, pos = parse_sed('sed.txt', position=True) #reads the file
bag = get_bag(pos) #Returns bag if sed was part of training
bins = bin_data(sed)

# estimate peak frequency
estimator = PeakFrequencyEstimator()
peak, err = estimator(bins, bag)
# estimate peak flux
estimator = PeakFluxEstimator()
peak, err = estimator(bins, bag)

References

• Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles, Lakshminarayanan et. al., 2016
• The Fourth Catalog of Active Galactic Nuclei Detected by the Fermi Large Area Telescope, The Fermit-LAT collaboration, 2019
• The 3HSP catalogue of Extreme & High Synchrotron Peaked Blazars, Chang et. al., 2019
• The 5th edition of the Roma-BZCAT. A short presentation, Massaro et. al., 2015