Neural Network identification of Dark Star Candidates. I. Photometry

TL;DR

The paper develops feed-forward neural networks to identify photometric Dark Star candidates in JWST/JADES data and to estimate their masses and redshifts under two DM heating scenarios: adiabatic contraction and DM capture. It trains on 10,000 TLUSTY-based simulated spectra across 14 NIRCam bands with realistic noise, achieving high fidelity in parameter recovery and enabling rapid candidate identification. Applied to real data, the method confirms two previously proposed candidates and discovers six new photometric candidates, delivering massive speed-ups over prior $\chi^2$-based fitting. This work demonstrates the viability of ML-driven photometric searches for Dark Stars in large JWST datasets and sets the stage for automated, large-scale discovery with spectroscopic follow-up to validate candidates.

Abstract

The formation of the first stars in the universe could be significantly impacted by the effects of Dark Matter (DM). Namely, if DM is in the form of Weakly Interacting Massive Particles (WIMPs), it could lead to the formation (at $z\sim 25-10$) of stars that are powered by DM annihilations alone, i.e. Dark Stars (DSs). Those objects can grow to become supermassive ($M\sim 10^6 \Msun$) and shine as bright as a galaxy ($L\sim 10^8 \Msun)$. Using a simple $χ^2$ minimization, the first three DSs photometric candidates (i.e. \JADESeleven, \JADEStwelve, and \JADESzthirteen) were identified by \cite{Ilie:2023JADES}. Our goal is to develop tools to streamline the identification of such candidates within the rather large publicly available high redshift JWST data sets. We present here the key first step in achieving this goal: the development and implementation of a feed-forward neural network (FFNN) search for Dark Star candidates, using data from the JWST Advanced Deep Extragalactic Survey (JADES) photometric catalog. Our method reconfirms JADES-GS-z13 and JADES-GS-z11 as dark star candidates, based on the chi-squared goodness of fit test, yet they are $\sim10^4$ times faster than the Neadler-Mead $χ^2$ minimization method used in \cite{Ilie:2023JADES}. We further identify six {\it new photometric} Dark Star candidates across redshifts $z \sim 9$ to $z \sim 14$. These findings underscore the power of neural networks in modeling non-linear relationships and efficiently analyzing large-scale photometric surveys, advancing the search for Dark Stars.

Figures (6)

• Figure 1: Schematic of the feed-forward neural network (FFNN) used in this work. The input consists of 14 or 11 photometric bands from JWST/NIRCam observations, depending on data availability. The network consists of fully connected hidden layers with ReLU activations and predicts two output parameters: stellar mass (in units of $10^5 M_\odot$) and redshift ($z$). Separate models are trained for adiabatic contraction and capture SMDSs formation scenarios.
• Figure 2: Training and validation loss, according to equation \ref{['eq:Loss']}, for all four neural networks. Panel a: Adiabatic contraction model with 14 inputs. Panel b: Adiabatic contraction model with 11 inputs. Panel c: Capture case model with 14 inputs. Panel d: Capture case model with 11 inputs. All models exhibit consistent convergence, with training and validation loss stabilizing after approximately 50 epochs. The small gap between training and validation loss indicates effective generalization to unseen data.
• Figure 3: Predicted versus true values of stellar mass and photometric redshift for the adiabatic contraction SMDSs models. Panel a: Model with 14 inputs, achieving $R^2 = 0.995$ for mass and $R^2 = 0.989$ for redshift. Panel b: Model with 11 inputs, achieving $R^2 = 0.994$ for mass and $R^2 = 0.945$ for redshift. The close alignment of points along the $y=x$ line reflects the high accuracy of the models in predicting both mass and redshift.
• Figure 4: Predicted versus true values of stellar mass and photometric redshift for the DM Capture SMDSs models. Panel a: Model with 14 inputs, achieving $R^2 = 0.996$ for mass and $R^2 = 0.989$ for redshift. Panel b: Model with 11 inputs, achieving $R^2 = 0.993$ for mass and $R^2 = 0.963$ for redshift.
• Figure 5: Comparison of predicted photometric flux density distributions to NIRCam data for JADES-GS-z11 and JADES-GS-z13, SMDSs photometric candidates first identified in Ilie:2023JADES. The solid blue lines represent the FFNN predicted SEDs, with the red boxes representing average flux density in each photometric filter considered. The black points with error bars indicate JWST observations. Left: JADES-GS-z11 with $z=11.54$, $\mu=1$, $\chi^2=10.69$, and predicted stellar mass $7.20 \times 10^5 M_\odot$ using the adiabatic contraction (AC) model. Right: JADES-GS-z13 with $z=12.85$, $\mu=1$, $\chi^2=20.29$, and predicted stellar mass $1.65 \times 10^6 M_\odot$ using the capture case model. The throughput curves of the photometric bands are shown below for reference, demonstrating the wavelength range contributing to each band.