Repository files navigation

We at DeepLense explore cutting-edge Machine Learning techniques for the study of Strong Gravitational Lensing and Dark Matter Sub-structure. We use both simulated and real lensing images, for a variety of tasks, using a variety of techniques.

We also actively mentorGoogle Summer of Code (GSoC) projects, that you can find listedhere.

1. Find below a description ofgravitational lensing anddark matter sub-structure.
2. Section 2 contains a detailed description of the datasets used in the various projects
3. Section 3 beins with a short description followed by an expansion on the various (GSoC) projects conducted at DeepLense

Gravitational lensing is the phenomenon of the bending of light in the gravity of a massive celestial object (such as a massive galaxy or a group of galaxies); the object essentially behaving as a cosmic lens. We, as a result see the distorted image(s) of light sources (typically another galaxy) behind it.

The dynamics of lensing depends on both the composition of the lens and the nature of the source. We explore different lens models and source light profiles including real galaxy images in DeepLense, that you can findhere.

In DeepLense, we’re mainly dealing with three kinds of simulated Dark Matter:

• Axion Dark Matter (Vortex): Axions are hypothetical particles that are considered as candidates for dark matter. In the context of axion dark matter, vortex substructures refer to specific topological features that can form in the distribution of axion fields.
• Cold Dark Matter (Subhalo): This model suggests that dark matter consists of slow-moving particles. In the CDM paradigm, smaller clusters of dark matter, known as subhalos, are approximated as “point masses.” This simplification facilitates computational modeling by treating these subhalos as singular points in the overall distribution of dark matter.
• No-Substructure Dark Matter: Unlike the CDM model, the “no-substructure” approach assumes that dark matter is evenly spread out, devoid of any smaller-scale clusters or sub-halos. This stands in stark contrast to the hierarchical structuring and layering of sub-halos within larger halos as predicted by CDM models.

All datasets are constructed using Lenstronomy, by Michael W. Toomey, as presented in theirrepository.

Classification of lensing images into their intrinsic dark matter sub-structure can reveal information about their characteristics and ways to study them. DeepLense presents an extensive array of techniques to study dark matter substructure involving both the simulated and real galaxy datasets.

Archil Srivastava, as part of theirGSoC 2022 project, explores the potency of variants of Vision Transformers (EfficientNet, ViT, ConViT, CrossViT, Bottleneck Transformers, EfficientFormer, CoaT, CoAtNet and Swin) on the classification of dark matter substructure on the three datasets, Model 1, 2 and 3.

Mriganka Nath employs unsupervised domain adaptation techniques as part of theirGSoC 2022 project such as ADDA, AdaMatch and self-ensembling from simulated lensing images to classify dark matter substructure of real lensing images from the Hyper Suprime-Cam (HSC) Subaru Strategic Program Public Data Release 3.

Lucas Jose bridges the gap between relativistic Physics principles and machine learning models in theirGSoC 2023 project, where they develop a Physics-Informed transformer architecture that outperforms its traditional counterparts.

Ashutosh Ojha builds on to improve Lucas' Physics-Informed transformer in theirGSoC 2024 project, through the inclusion of Physics-Informed pre-processing, and a GradCAM visualization allowing for the interpretation of its working.

Sreehari Iyer evaluates Transformer-based self-supervised learning techniques through theirGSoC 2024 project, utilizing the real-world strong gravitational lensing dataset, Model 4. ViT-S and ViT-B are selected as the backbone and SimSiam, DINO and iBOT for self supervised training.

Yashwardhan Deshmukh compares the performance of the self-supervised learning techniques in theirGSoC 2023 project, Contranstive Learning and Bootstrap Your Own Latent (BYOL) on the three datasets, Model 1, 2 and 3.

Marcos Tidball performs unsupervised domain adaptation in theirGSoC 2021 project, to mitigate the poor generalization of models trained on simulated data to real lensing images using ADDA, Self-Ensemble, CGDM and AdaMatch.Their work has been published as apaper in the Astrophysical Journal.

Apoorva Singh, in theirGSoC 2021 project andGeo Jolly, in theirGSoC 2023 project exploit the inherent symmetries present in the strong lensing system (such as rotations and reflections) using equivariant neural networks, to extract dark matter sub-structural information.

Kartik Sachdev, through theirGSoC 2022 andGSoC 2023 projects performs benchmarking of an extensive variety of vision transformer architectures, and contrasting of ten supervised and two self-supervised learning frameworks on the classification of dark matter substructure. They use the three datasets, Model 1, 2 and 3.

Saranga K Mahanta, in theirGSoC 2022 project conducts a study on strong lensing through a variety of tasks on strong lensing images including classification, anomaly detection and regression using several different neural network architectures, on the three datasets, Model 1, 2 and 3.

Another means of dark matter study through strong lensing is through the approximation of their properties.Yurii Halychanskyi andZhongchao Guan approximate the mass density of vortex substructure of dark matter condensates on the three datasets, Model 1, 2 and 3.Yurii uses the ResNet18Hybrid and CmtTi architectures in theirGSoc 2021 and2022 projects, while Zhongchao demonstres with ResNet18, ViT, CNN-T, MobileNet V2 and CvT-13, in theirGSoc 2022 project.

Finally, DeepLense help combat the problem of noisy and low-resolution of real lensing images through various super-resolution techniques. Denoising and upscaling of lensing images can help make their study more accurate.

Atal Gupta achieves super-resolution of the real-galaxy lensing dataset, in theirGSoC 2024 project, Model 4 using a variety of Diffusion Models (DDPM, SR3, SRDiff, ResShift and CG-DPM).

Pranath Reddy performs a comparative study of the super-resolution of strong lensing images in theirGSoC 2023 project, using Residual Models with Content Loss and Conditional Diffusion Models, on the Model 1 dataset.

Anirudh Shankar explores the unsupervised super-resolution of strong lensing images through a Physics-Informed approach in hisGSoC 2024 project, built to handle sparse datasets. They use custom datasets using different lens models and light profiles.