3D extinction mapping of the Milky Way using Convolutional Neural Networks: Presentation of the method and demonstration in the Carina Arm region (2201.05571v1)

Abstract: Context. Several methods have been proposed to build 3D extinction maps of the Milky Way (MW), most often based on Bayesian approaches. Although some studies employed ML methods in part of their procedure, or to specific targets, no 3D extinction map of a large volume of the MW solely based on a Neural Network method has been reported so far. Aims. We aim to apply deep learning as a solution to build 3D extinction maps of the MW. Methods. We built a convolutional neural network (CNN) using the CIANNA framework, and trained it with synthetic 2MASS data. We used the Besan\c{c}on Galaxy model to generate mock star catalogs, and 1D Gaussian random fields to simulate the extinction profiles. From these data we computed color-magnitude diagrams (CMDs) to train the network, using the corresponding extinction profiles as targets. A forward pass with observed 2MASS CMDs provided extinction profile estimates for a grid of lines of sight. Results. We trained our network with data simulating lines of sight in the area of the Carina spiral arm tangent and obtained a 3D extinction map for a large sector in this region ($l = 257 - 303$ deg, $|b| \le 5$ deg), with distance and angular resolutions of $100$ pc and $30$ arcmin, respectively, and reaching up to $\sim 10$ kpc. Although each sightline is computed independently in the forward phase, the so-called fingers-of-God artifacts are weaker than in many other 3D extinction maps. We found that our CNN was efficient in taking advantage of redundancy across lines of sight, enabling us to train it with only 9 sightlines simultaneously to build the whole map. Conclusions. We found deep learning to be a reliable approach to produce 3D extinction maps from large surveys. With this methodology, we expect to easily combine heterogeneous surveys without cross-matching, and therefore to exploit several surveys in a complementary fashion.

• The paper introduces a CNN-based framework to map 3D extinction using color-magnitude diagrams from simulated 2MASS-like star catalogs.
• It achieves resolutions of 100 pc and 30 arcminutes across distances up to 10 kpc, efficiently integrating heterogeneous survey data.
• The approach outperforms traditional Bayesian and MCMC methods, reducing artifacts and accurately detecting distal interstellar structures.

3D Extinction Mapping of the Milky Way using Convolutional Neural Networks

The paper "3D extinction mapping of the Milky Way using Convolutional Neural Networks" presents an innovative approach to constructing three-dimensional extinction maps of the Milky Way, particularly applied to the Carina Arm region. The research extends the capabilities of deep learning, specifically Convolutional Neural Networks (CNNs), to derive extinction profiles across large volumes of the Milky Way, demonstrating its utility compared to traditional Bayesian techniques.

The authors trained a CNN using the CIANNA framework with synthetic data modeled after the 2MASS catalog. This involved generating mock star catalogs using the Besançon Galaxy model and simulating 1D Gaussian random field extinction profiles. The colored and magnitude data from these simulations were used to develop color-magnitude diagrams (CMDs) that served as the training grounds for the CNN. Remarkably, the trained network was able to process the convolutional input and efficiently produce a coherent and high-resolution 3D extinction map despite the "fingers-of-God" effect being less pronounced than in traditional 3D extinction maps. This performance was notable, given that training effectively leveraged redundancy across different sightlines by employing only nine lines of sight simultaneously.

The resulting 3D map for a significant section of the Carina Arm region showcases distance and angular resolutions of 100 pc and 30 arcminutes, respectively, with reach extending up to roughly 10 kpc. The ability of this CNN-driven approach to simultaneously process and combine heterogeneous survey data without requiring cross-matching is an impressive advantage over conventional methods, which often demand careful catalog alignment and are limited by spectral domain constraints. This facilitates its application in integrating data from multiple surveys such as Gaia and PAN-STARRS 1 with greater efficiency and coherence in inferred interstellar structures.

Empirically, the method supported the identification of dense interstellar medium structures within the Carina region by achieving consistent alignment with known astrophysical properties and regions like HII complexes and giant molecular clouds. The comparisons drawn between this method's results and those from MCMC approaches reveal substantial advantages in detecting distal structures and producing artifacts with reduced exaggeration.

The CNN framework set forth in this paper paves the way for future advances in Galactic mapping. Its capability to synthesize observations from a range of data sources provides meaningful flexibility for future theoretical inquiries and observational campaigns. However, challenges such as sensitivity to the realism in training datasets and the integration of input reference CMDs highlight areas for further refinement and development of more adaptive algorithms.

Looking forward, this research implies a significant shift in the methodology for large-scale Galactic surveys. Embracing such deep learning techniques could enhance our understanding of Galactic structure and extinction while efficiently bridging data from multiple astrophysical surveys without the conventional methodological constraints. Further research could explore the flexibility this approach offers in integrating database enhancements from ongoing space and ground-based observatories, potentially leading to a more cohesive and detailed picture of the Milky Way's interstellar medium.