Machine Learning for Astronomy: Galaxy Spectra going to the Gym

Title: Deep learning prediction of galaxy stellar populations in the low-redshift Universe

Authors: Li-Li Wang, et al.

First Author’s Institution: Dezhou University

Status: Published in MNRAS [open access]

A galaxy’s spectrum is a host of incredible information, if only we as scientists are capable of interpreting it. It can tell us hints about the mass of the galaxy, the age of the galaxy, and perhaps most importantly, even the kinds of stars that are present in the galaxy. Once you observe the spectrum of light from a galaxy, it becomes a game of interpretation. But how do we get from photon spikes on a detector to all this information?

The idea is to match this observed spectrum to some modeled spectrum where you prescribe galaxy ‘properties’ like the specific age, mass, and metallicity of the galaxy and the stars in it. The closer the observed spectrum looks to the modeled spectrum, the more you can be sure the observed galaxy has the same properties as the model.

Of course, this is done with quite rigorous statistics. Typically, an astronomer would use what’s called a full spectrum fitting method to find the stellar population properties of the observed galaxy. This uses a Chi-squared fitting routine (or something like MCMC) to go through a wide range of possible properties and determine what mixture best matches the observed spectrum by minimizing the residuals. This, however, is quite time-intensive! It takes quite a while for a computer to go through every possible parameter combination to find what is closest, and if you have a lot of parameters you’re testing (like, over 2) this can take several hours for a large sample of galaxies.

Luckily, there are researchers who are hoping to help cut down on this time so our predictions can get more refined. The authors of today’s bite try their hand at using convolutional neural networks (CNNs) to predict properties of the stellar population in galaxies.

Getting the spectra in the gym

In their analysis, they use a sample of 100,000 randomly-selected low-redshift (0.002 < z < 0.3) SDSS galaxies with a signal to noise ratio above 5. To train the network, they get the ‘true’ properties of the sample from the more common full spectrum fitting routine, specifically using the code PPXF. In their analysis, they attempt to predict four properties: age, metallicity, reddening (E(B-V)), and velocity distribution of the stars.

With these ‘true’ property values, they are able to train their neural network. The idea is for the neutral network to minimize what is called the loss function, which is just the mean-squared error of the predicted and true values. Figure 1 shows the active minimization of the loss function, which eventually converges to a stable solution. This means the neural network has gotten as close as it can to the ‘true’ values!

Figure 1. Loss function of the neutral network. It is seen to decrease almost exponentially with time (epoch), reaching a steady state after a while. Figure 3 in original paper

Training season’s over

Notably, it took only about two hours to train the model on this dataset and to estimate the parameter values, in contrast to the 20 hours it took to use PPXF. The CNN was able to predict the true values of the parameters with good accuracy. The larger scatter in the velocity dispersion is typical of traditional fitting methods.

Figure 2. Parameter values predicted by CNN vs. true values from PPXF for each property: 1. Age 2. Metallicity 3. E(B-V) 4. VD. Blue line shows a one-to-one relationship. Figure 4 in original paper

Finally, the authors were interested in exploring how the CNN’s accuracy could be affected by any data biases. They first explored the question of data quality, to see if the model was more accurate on higher signal-to-noise data. They found that galaxy spectra with a lower signal-to-noise ratio had only marginally worse property prediction in their CNN. They then asked the same question of redshift, exploring if the CNN was biased to any specific range. They found that the CNN was worse at property prediction at the ‘extrema,’ so when the redshift was either much higher than the rest or much lower. Finally, they considered any bias in regard to the type of galaxy, splitting the data into the categories of “passive” galaxies, “star-forming” galaxies, and “composite” galaxies. They found that the CNN gave the best property predictions for passive galaxies, and the worst for star-forming galaxies.

New machine learning methods are becoming more common in astronomy research, but much more training is needed to be able to apply these methods ubiquitously. But the prospect looks bright for these methods and their ability to save some precious time.

Astrobite edited by Tori Bonidie

Featured image credit: Nilda from Wikipedia, Public Domain

About Caroline von Raesfeld

I'm a second-year PhD student at Northwestern University. My research explores how we can better understand high-redshift galaxy spectra using observations and modeling. In my free time, I love to read, write, and learn about history.

Discover more from astrobites

Subscribe to get the latest posts to your email.

Leave a Reply