Neural Network identification of Dark Star Candidates. II. Spectroscopy
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.
Vera: We're looking at the paper: "Neural Network identification of Dark Star Candidates. II. Spectroscopy".
Jocelyn: Here's the gist of the paper. Neural Network identification of Dark Star Candidates.
Vera: II.
Jocelyn: Spectroscopy is a study that develops and validates a Feedforward Neural Network (FFNN) to identify spectroscopic dark star candidates in JWST NIRSpec data, establishing a robust and efficient tool for probing Dark Stars and understanding their role in cosmic evolution.
Vera: Neural Networks (NNs) have experienced an unprecedented surge in applications across diverse fields, leveraging the power of pattern recognition to address complex real-world challenges.
Paper summary: Vera: So we're diving into this paper now, "Neural Network identification of Dark Star Candidates. II. Spectroscopy." Essentially, the authors are proposing a feedforward neural network to identify spectroscopic dark star candidates using data from JWST's NIRSpec instrument.
Jocelyn: The core thesis is that these neural networks can model the highly non-linear relationships in spectral data to infer crucial parameters like mass and redshift of these dark stars.
Subrahmanyan: They are specifically looking at dark stars, which are objects whose power might come from dark matter annihilations instead of traditional nuclear fusion, offering a unique window into dark matter physics.
Vera: The paper claims they developed this feedforward neural network and trained it on synthetic spectra of Supermassive Dark Stars to identify candidates in real JWST data.
Jocelyn: They are using this model to predict stellar mass and redshift from input spectra, and they want to use it as a tool for probing dark stars in the early universe.
Subrahmanyan: The significance is that this machine learning approach is being used to tackle complex problems involving noisy and incomplete datasets, which is essential for high-redshift studies.
Vera: The authors are showing that this neural network can identify these candidates with R squared scores greater than zero point nine nine for both the mass and redshift predictions they make.
Jocelyn: They also emphasized that the model's ability to fit the continuum of the spectra is quite remarkable when compared to their models.
Subrahmanyan: The real payoff here, from a cosmological standpoint, is creating a pipeline capable of automatically and efficiently handling massive amounts of high-redshift universe archival JWST spectroscopic data.
Vera: So it's about using this neural network to create an automatic identification system for these dark star candidates within big astronomical datasets.
Jocelyn: This work sets up a pathway for integrating advanced machine learning models into workflows that can process the kind of massive observational data that JWST is generating.
Subrahmanyan: It shows how ML can be used to extract meaningful information from complex astrophysical signals, which is vital for understanding structure formation in the early universe.
Conclusion: Vera: Thinking about this paper, "Neural Network identification of Dark Star Candidates. II. Spectroscopy," the authors are essentially showing how to use a neural network for spectroscopic dark star candidate identification with JWST data.
Jocelyn: The title itself points to the specific technique—using a feedforward neural network—to tackle this particular astrophysical problem of finding dark stars spectroscopically.
Subrahmanyan: The implications are that we might gain a new observational method for studying dark matter powered stars by leveraging these automated machine learning tools.
Vera: It moves us toward creating an automatic pipeline where researchers can efficiently sift through huge amounts of archival JWST data to find these candidates without having to manually analyze every single spectrum.
Jocelyn: It suggests that we need to look closely at how observational uncertainties are handled, since the authors noted that the model doesn't fully incorporate all of them yet.
Subrahmanyan: That means future work needs to focus on making these models more rigorous by using Bayesian neural networks so they can produce probability distributions for mass and redshift predictions.
Vera: So, it's about moving from just having a good prediction to having a system that can quantify the uncertainty of that prediction.
Jocelyn: And ultimately, the goal is to build tools capable of handling the sheer scale of data coming from telescopes like JWST automatically for dark star research.
Adiba Amira Siddiqa, Sayed Shafaat Mahmud, Cosmin Ilie
Bryn Mawr College · Colgate University
astro-ph.IM, astro-ph.CO
Submitted: 2025-11-06
Updated: 2025-11-06
Comments: 9 pages, 4 figures, submitted to the Astrophysical Journal
Journal ref: Astronomy and Computing, Volume 58, Article 101195 (2027)
DOI: 10.1016/j.ascom.2026.101195
Project page: https://ceers.github.io/Freese
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 80/100
The gist: Neural Network identification of Dark Star Candidates.
Key concepts
- Feedforward Neural Network (FFNN)
- A type of neural network where information moves in one direction, from the input layer through hidden layers to the output layer. It uses interconnected nodes to learn complex patterns from data, making it suitable for predicting properties like mass and redshift from spectral data.
- Spectroscopic Dark Star Candidates
- These are hypothetical stars that are thought to be dark because they emit very little light. The study uses spectroscopy—analyzing the light spectrum—to find evidence of these stars in JWST data, helping scientists understand dark matter powered stars.
- Adam Optimization Algorithm
- An efficient method used during the training of neural networks to adjust the network's internal parameters. It helps the FFNN learn how to best map input spectra to physical properties like mass and redshift, ensuring accurate predictions.
- R2 Score
- A statistical measure indicating how well a model fits the data. An R2 score of 0.99 or higher means the neural network's predictions for a dark star's mass and redshift are extremely close to the actual, true values.
Terminology
Summary
Neural Network identification of Dark Star Candidates. p. 0 Draft version November 7, 2025
Typeset using LATEX default style in AASTeX631
Neural Network identification of Dark Star Candidates. Spectroscopy
Adiba Amira Siddiqa,1 Sayed Shafaat Mahmud,2 and Cosmin Ilie2
1Bryn Mawr College
101 N Merion Ave
Bryn Mawr, PA 19010, USA
2Colgate University
13 Oak Drive
Hamilton, NY 13346, USA
ABSTRACT
Some of the first stars in the Universe might be powered by Dark Matter (DM) annihilations, rather
than nuclear fusion. With this in mind, the FFNN model we developed and validated
in this work will be adapted for Bayesian uncertainty analyses and automatic analyses of NIRSpec
publicly available data for high redshift objects. This study establishes a robust and efficient tool for
probing Dark Stars and understanding their role in cosmic evolution.Where the paper says it
Where the paper says it
Introduction to Machine Learning in Astrophysics
Neural Networks (NNs) have experienced an unprecedented surge in applications across diverse fields, leveraging the power of pattern recognition to address complex real-world challenges. Deep learning has revolutionized domains such as image classification, text classification, speech recognition, genome research, and astrophysics. p. 0 2022;
Mienye & Swart 2024). Deep learning, a specialized branch of machine learning, has revolutionized domains such as
image classification (Wang et al. 2025a; Zeraatgari et al. The application of machine learning (ML) in astrophysics has grown rapidly over the past decade, transforming
how researchers analyze large and complex datasets. 2
techniques, Feed Forward Neural Networks (FFNNs) have emerged as a cornerstone in astrophysics due to their ability
to model highly non-linear relationships, enabling applications such as photometric redshift estimation (Razim et al. 2021; Curran et al.Where the paper says it
Where the paper says it
Where the paper says it
Model Architecture and Training Protocol
The FFNN developed in this work is designed to infer crucial parameters such as mass and redshift of Dark Stars that can fit observed spectra taken with NIRSpec. The network architecture includes an input layer followed by three fully connected hidden layers containing 256, 128, and 64 neurons, respectively. p. 1 2 microns, as appropriate based on the sensitivity range of NIRSpec. The network architecture includes an input layer followed by three fully connected hidden layers containing 256, 128,
and 64 neurons, respectively. The network architecture includes an input layer followed by three fully connected hidden layers containing 256, 128,
and 64 neurons, respectively. Each hidden layer applies a rectified linear unit (ReLU) activation function (Agarap
2018), which introduces non-linearity and allows the network to learn more complex mappings between inputs and
outputs. Each hidden layer applies a rectified linear unit (ReLU) activation function (Agarap
2018), which introduces non-linearity and allows the network to learn more complex mappings between inputs and
outputs. The final output layer has two neurons, one for mass and one for redshift, both treated as continuous
variables. This introduces variability that helps the network generalize to noisy observations. We train the network
using the Adam optimization algorithm (Kingma & Ba 2014), which adjusts learning rates adaptively to accelerate
convergence and improve stability. For the capture-based models, we use a learning rate of 5 × 10−4
; for adiabatic
contraction models, we use 1×10−3
. The loss function (Equation 1) is the weighted average of the mean squared error
(MSE), which quantifies the sum of average squared difference between predicted and true values of dark star mass
and redshift.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Performance Validation and Comparison
The model performance on test datasets showed strong agreement with the true values for both SMDS mass and redshift, achieving R2 scores of ≳ 0.99. The FFNN spectral fits demonstrated remarkable agreement between data and models at the level of the continuum. p. 4 4 depicts a comparison of
the the model-predicted spectra (orange line) with the observed data (blue line) and its associated uncertainty (gray
shaded regions). Based on the normalized residuals (depicted with black in the sliver rectangular plots underneath each
spectral fit) we conclude that our FFNN spectral fits demonstrate remarkable agreement between data and models, This procedure
typically requires hours of computation for a single object. By contrast, our neural network predicts stellar mass and
redshift in milliseconds—about 105
times faster.Where the paper says it
Where the paper says it
Limitations and Future Work
Several limitations remain, including the possible role of nebular emission from Dark Stars which will be taken into account in future iterations. Furthermore, the model does not fully incorporate observational uncertainty into its predictions. p. 5 First, the possible role of nebular
emission from Dark Stars will be taken into account in future iterations of our FFNN algorithm, such that we can
also couple it to the morphological analysis pipeline described in Ilie et al. 1 Moreover, the spectra used in
this study are accompanied by associated errors, but our model does not fully incorporate this uncertainty into its
predictions. To better capture how uncertainties in dark star spectra affect the inferred masses and redshifts, and to
reflect the probabilistic character of neural networks, future work will investigate Bayesian Neural Networks (BNNs). A Bayesian framework would allow us to quantify uncertainties more rigorously by producing posterior distributions
for the predicted parameters, rather than point estimates. For instance, embedding dark stars within a host galaxy and
examining how the surrounding stellar and interstellar environments modify the spectrum could yield more realistic
models. Moreover, all SMDSs models used so far in the literature to identify Dark Star candidates were based
100 GeV WIMPs, annihilating with the canonical cross section that leads to the observed relic abundance of DM:
⟨σv⟩ ∼ 3 × 10−26cm3
s
−1
. Moreover, no MIRI data was
yet used in those analyzes. So, in order to further test the Dark Star hypothesis, we will create a pipeline based on the
extremely fast, yet robust FFNNs described this work (and Paper I) that is able to automatically and efficiently handle
the large amounts of high-redshift Universe archival JWST spectroscopic (and photometric) data available either in the
MAST catalog (STScI 2023), or from dedicated survey websites such as: CEERS (Finkelstein & CEERS Collaboration
2025), JADES (JADES Collaboration 2025), GLASS (GLASS-JWST Team 2023), UNCOVER (UNCOVER Team
2024), and PRIMER (PRIMER Collaboration 2023).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion and Impact
The confirmation of even a single Supermassive Dark Star, based on this or potentially other distinct spectroscopic signatures, would have enormous impact on astronomy, astrophysics, and particle physics. The study provides a foundation for integrating advanced machine learning models in an automatic pipeline to identify Dark Star candidates within large amounts of archival JWST data. p. 1 2025b) provide a foundation for integrating advanced
machine learning models in an automatic pipeline we plan to design with the purpose identifying Dark Star candidates
within large amount of archival JWST data (spectroscopic and photometric) from deep surveys dedicated to probing
the high redshift universe. Neural Network Architecture
As mentioned in the previous section, in this work we develop a feedforward neural network (FFNN) in order
to identify Supermassive Dark Star candidates based on observed JWST NIRSPec spectra for high-z objects, and
estimate their Mass (M) and redshift (z). The confirmation of even a single
Supermassive Dark Star, based on this, or potentially other, distinct spectroscopic signatures, would have enormous
impact on astronomy, astrophysics, and particle physics. It would indirectly verify the existence of self-annihilating
DM particles, while, at the same time, opening up a new field of Astronomy and Astrophysics: the observational study
of Dark Matter powered stars.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Automated spectroscopic candidate identification for Supermassive Dark Stars (SMDS) can be performed in milliseconds, over 10,000 times faster than traditional Monte Carlo parameter estimation methods used previously.
-
The FFNN model can predict crucial physical parameters such as mass and redshift of Dark Stars from JWST NIRSpec spectra, allowing for rapid screening of large amounts of high-redshift data.
-
The system can be adapted for Bayesian uncertainty analyses, producing posterior distributions for predicted parameters instead of just point estimates to rigorously quantify uncertainties in inferred masses and redshifts.
Abstract
Some of the first stars in the Universe might be powered by Dark Matter (DM) annihilations, rather than nuclear fusion. Those objects, i.e. Dark stars (DS), offer a unique window into understanding DM via the observational study of the formation and evolution of the first stars and their Black Hole (BH) remnants. In (Paper I) we introduced a feedforward neural network (FFNN) trained on synthetic DS photometry in order to detect and characterize dark star photometric candidates in the early universe based on data taken with the NIRCam instrument onboard the James Webb Space Telescope (JWST). In this work we develop a FFNN trained on synthetic DS spectra in order to identify spectroscopic dark star candidates in the data taken with JWST's NIRSpec instrument. In order to validate our FFNN model we apply it to real data for the four spectroscopic Supermassive Dark Star (SMDS) candidates recently identified in and reconfirm that indeed,,, and have spectra that are consistent with those of Supermassive Dark Stars. The main advantage of our FFNN model, in comparison to the Nedleaer-Mead Monte Carlo parameter estimator used in, is that the approach introduced here predicts parameters in milliseconds, over 10,000 times faster than the traditional method used in. With this in mind, the FFNN model we developed and validated in this work will be adapted for Bayesian uncertainty analyses and automatic analyses of NIRSpec publicly available data for high redshift objects. This study establishes a robust and efficient tool for probing Dark Stars and understanding their role in cosmic evolution.
Sources
- Deep Learning using Rectified Linear Units (ReLU)
- Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture
- Multi-column Deep Neural Networks for Image Classification
- A brief introductory guide to TLUSTY and SYNSPEC
- Adam: A Method for Stochastic Optimization
- VADER: A Variational Autoencoder to Infer Planetary Masses and Gas-Dust Disk Properties Around Young Stars
- DINOv2: Learning Robust Visual Features without Supervision
- PyTorch: An Imperative Style, High-Performance Deep Learning Library
- Explainable Artificial Intelligence: Understanding, Visualizing and Interpreting Deep Learning Models
- Text Classification via Large Language Models
- Deep Learning for Genomics: A Concise Overview
Related papers
- On-sky demonstration of a vector Zernike wavefront sensor in a cascaded adaptive optics system
- Euclid Data Release 1 (DR1): Calibration strategy and observations
- The General Antiparticle Spectrometer (GAPS) Antarctic Balloon Payload
- Detection and excision of radio-frequency interference in radio arrays
- Unified matched-filter and excess-power pipeline for gravitational-wave detector characterization
- Ushering in sub-arcsecond angular resolution observatories in the far-infrared: The case for a double-Fourier interferometer