Research Projects
Solving PDEs using Physics Informed Neural Networks (PINNs)
Generative modelling using Generative Adversarial Networks (GANs), cGANs, UNET, UNET+ ATTENTION and Diffusion models
Detecting Exoplanet Mass using Computer Vision
(CNN and MLP models)
Uncertainty Estimation using Bayesian CNNs
We develop GRINN, PINN-based code, to solve coupled time-dependent PDEs for specific initial and boundary conditions to simulate complex hydrodynamic systems. We simulate the non-linear gravitational instability by solving the hydrodynamic equations using GRINN. In 3D, GRINN solvers are 10 times more efficient compared to the Finite Difference schemes.
Schematic of the PINN-based GRINN workflow.
Animations showing the evolution of the initial sinusoidal perturbation subject to different initial conditions
GAN: We train a Generative Adversarial Network(GAN) to generate face-on images of protoplanetary disks (PPDs) from arbitrarily oriented input images. This is crucial in studying protoplanetary disks since most observed disks are oriented at a random angle in the sky. Hence it is challenging to study such disks since they are partially visible. With the aid of Generative models, it is impossible to generate face-on images instantly, enabling astronomers to constrain disk properties better and understand the fundamentals of Planet formation.
Generative MODEL to rotate the protoplanetary disk images from any arbitrary orientation to face-on images: We adopt the UNET-based PIX2PIX code: The image demonstrates an example case, where the GAN model rotates the input image to generate a face-on image.
​
​
ATTENTION-GAN: We implement an attention module to the UNET-based GAN model. We wanted to explore if the attention block along with the skip connection in the UNET-based generator improves the generative capacity. The project is still on going
Detecting Exoplanet in their early stage of formation is difficult since they are mostly hidden within dusty protoplanetary disks(PPDs). Traditional techniques require expensive computation and modeling to find and characterize this unseen exoplanet.
We trained a Convolutional Neural Network (CNN, here specifically ResNet50) to detect these hidden planets using computer vision directly from the images of the PPDs. These models are trained to identify features associated with planet formation (even when the planet itself is not visible) and thus learn the underlying physics. Once trained we predict exoplanet masses directly from images of protoplanetary disks hosting a planet cutting down computation times from weeks to seconds.
Code: https://github.com/sauddy/DPNNet-2.0
https://github.com/sauddy/DPNNet-RT
Papers:
1. DPNNet - https://iopscience.iop.org/article/10.3847/1538-4357/aba95d/meta
2. DPNNet-2.0 - https://iopscience.iop.org/article/10.3847/1538-4357/ac1518/meta
​
​
More details can be found in my blog post.
We introduce the use of Bayesian Formalism in Neural Networks. We develop a Bayesian deep-learning network, "DPNNet-Bayesian," which can predict planet mass from disk gaps and also provides the uncertainties associated with the prediction. A unique feature of our approach is that it can distinguish between the uncertainty associated with the deep-learning architecture and the uncertainty inherent in the input data due to measurement noise.
