Unveiling Hidden Lyman Alpha Emitters in the DESI DR1 Data

Introduction to the show: ident: Astrophysics Radio. Generated commentary on the latest astrophysics papers.

Vera: Hi, I'm Vera.

Jocelyn: And I'm Jocelyn, here with Subrahmanyan.

Vera: We're looking at the paper: "Unveiling Hidden Lyman Alpha Emitters in the DESI DR1 Data".

Jocelyn: Here's the gist of the paper. The gist We present an automatic method based on machine-learning convolutional neural network (CNN) architecture to detect Lyman alpha emitters (LAE) hidden in the Data Release 1 spectroscopic dataset of the Dark Energy Spectroscopic Instrument (DESI), which is crucial because current DESI pipelines are not designed to detect and measure the redshifts of galaxies at z > 2.

Vera: The CNN architecture consists of two parts: 1.

Jocelyn: The first part consists of fully-convolutional layers without any fully-connected layers, where each convolution layer is followed by a ReLU activation function and a max pooling layer to downscale the size of the input spectrum.

Vera: This part takes the blue-arm S/N spectrum (2438 pixels, 1 pixel = 0.8˚A) as input and outputs a lower resolution spectrum (76 pixels, 1 pixel ≈ 25˚A), where each output pixel value is from 0 to 1, indicating the probability of the corresponding region of the original spectrum containing a Lyα emission line.

Title and authors: Vera: They explain their approach by saying they first visually inspect thousands of DESI spectra to build a training set that has both LAEs and non-LAEs. That’s the truth table they use to teach their machine learning model what a real LAE looks like.

Jocelyn: Then, the core of it is this two-part CNN architecture. The first part takes the blue-arm S/N spectrum, which is two thousand four hundred thirty-eight pixels, and it outputs a lower resolution spectrum with seventy-six pixels where each pixel is just a probability between zero and one indicating if that region has a Ly alpha emission line.

Subrahmanyan: That downscaling process sounds like they’re trying to reduce the complexity of the input before the model tries to find those faint signals, which makes sense given how noisy these spectra can be.

Vera: And then they have this second part, which is a slicer and a fitter. The slicer picks the pixel from that first output with a probability greater than zero point five and then they extract that region of the original spectrum.

Jocelyn: After extracting that specific area, it goes to another part—a fitter—which uses two convolution layers followed by three fully connected layers to fit the red peak of the LAE Ly alpha emission line as a peak-normalized Gaussian.

Subrahmanyan: So they are essentially using a neural network to scan for the signal, isolate it, and then model its shape precisely, which is much more detailed than what previous pipeline tools can do.

The paper's summary: Vera: The authors point out a few things they improved over before. They simplified the architecture by removing fully-connected layers in the first part of the CNN, and they say this made it better because that specific type of architecture generalizes better for detecting Ly alpha lines.

Jocelyn: They also talk about balancing the training data, which is important because they added supplementary spectra like quasars and low-redshift ELGs to make sure the model doesn't just learn to predict the majority class, which could be non-LAEs.

Subrahmanyan: And then they have a post-processing step where they assume some identified Ly alpha features are actually other lines and calculate new "correct" redshifts based on that assumption, which helps remove sources with incorrect line identification, specifically removing two thousand seven hundred ninety-three sources with an error rate of zero point eight two percent.

Vera: They also introduce an empirical separation criterion between star-forming LAEs and AGN by using the CIV over Ly alpha flux ratio set at zero point one five as a threshold to help distinguish them.

Jocelyn: And looking ahead, they plan to use attention-based Transformer architectures, like those from AstroCLIP, to capture correlations between multiple lines like Ly alpha and CIV, which should allow the model to recover LAEs out to redshift three point five.

The paper's improvements: Vera: So in summary of "Unveiling Hidden Lyman Alpha Emitters in the DESI DR1 Data," they found a final catalog with nineteen thousand six hundred eighty-five LAEs detected, with a purity of ninety-five point two percent and a completeness of ninety-five point nine percent <ref:2605.12503#pg3> <ref:2605.12503#pg3>. These are sources mostly at the bright end of the luminosity function with Ly alpha luminosity LLy alpha greater than one hundred thousand erg per second at redshifts from two to three point five.

Jocelyn: That catalog is pretty substantial, and it shows that even with these limitations in the current DESI pipeline, you can extract a sample of sources that is mostly at the bright end of what you expect for those early galaxies.

Subrahmanyan: From a theoretical view, this suggests that the population of these LAEs dominating the fainter end is being systematically missed, so getting this sample helps us understand how structure forms at those high redshifts.

Vera: The team also found some interesting issues with the pipeline, like thirty-seven percent of LAEs didn't even have a robust redshift detection and about thirty-four percent were misidentified as quasars with correct redshifts <ref:2605.12503#pg1> <ref:2605.12503#pg1>.

Jocelyn: They also noted that another twenty one percent were misidentified as other lines like O II doublet lines and often got assigned very low redshifts, around zero point one to zero point five.

Subrahmanyan: Overall, this work provides a tool that helps us use the DESI DR1 data in a way it wasn't originally intended for, opening up new avenues for studying galaxy formation at high redshift.

Vera: This paper "Unveiling Hidden Lyman Alpha Emitters in the DESI DR1 Data" gives us a concrete catalog and some clear ways to improve how we process this kind of data before we move on to the next big thing.

Jocelyn: It's a solid step forward for making sure we don't miss these high redshift emitters when looking at surveys like DESI.

Subrahmanyan: The work shows that this method helps us use the DESI DR1 data in a way it wasn't originally intended for, opening up new avenues for studying galaxy formation at high redshift.

Conclusion: Vera: So we've been looking at "Unveiling Hidden Lyman Alpha Emitters in the DESI DR1 Data," and they've built a machine learning system to find Lyman alpha emitters hidden in the Dark Energy Spectroscopic Instrument data, specifically because current pipelines can't handle galaxies at redshift greater than two.

Jocelyn: Right. They used a convolutional neural network architecture to scan those spectra, and it works by first downscaling the input spectrum and then using a slicer and fitter to pinpoint the emission line itself.

Subrahmanyan: It’s about taking that raw spectral information that's hard to interpret with standard tools and turning it into a detectable signal for high-redshift sources.

Vera: The results they got are pretty solid, showing ninety-five point two percent purity and ninety-five point nine percent completeness on the test data. They found nineteen thousand six hundred eighty five LAEs between redshifts two and three point five in just twelve minutes on a single GPU.

Jocelyn: That catalog is pretty substantial, and it shows that even with these limitations in the current DESI pipeline, you can extract a sample of sources that is mostly at the bright end of what you expect for those early galaxies.

Subrahmanyan: From a theoretical view, this suggests that the population of these LAEs dominating the fainter end is being systematically missed by existing software, so getting this sample helps us understand how structure forms at those high redshifts.

Vera: They did point out some issues too, like thirty-seven percent of LAEs didn't even have a robust redshift detection and about thirty-four percent were misidentified as quasars with correct redshifts.

Jocelyn: And they noted another twenty one percent were misidentified as other lines like O II doublet lines and often got assigned very low redshifts, around zero point one to zero point five.

Subrahmanyan: Overall, this work provides a tool that helps us use the DESI DR1 data in a way it wasn't originally intended for, opening up new avenues for studying galaxy formation at high redshift.

Vera: This paper "Unveiling Hidden Lyman Alpha Emitters in the DESI DR1 Data" gives us a concrete catalog and some clear ways to improve how we process this kind of data before we move on to the next big thing.

Jocelyn: It's a solid step forward for making sure we don't miss these high redshift emitters when looking at surveys like DESI.

Subrahmanyan: We need more of this kind of method applied across different survey releases to get a better census of these early galaxies, and maybe look into those Transformer architectures they mentioned for capturing multiple lines together.