Flexible and Open to learn

Neural Networks, Disease Propagation Modelling, Physics-Informed models

The ultimate goal of my research is to model the dynamics of disease propagation by developing Algorithms which lie in the intersection of scientific machine learning (SciML) and epidemiological models. SciML tightly integrate scientific prior knowledge in the training of artificial neural networks. Disease propagation in epidemiology often takes the shape of dynamical systems defined by Partial Differential Equations (PDE). In this regard, the first attempt of my research has been to modify the physics-informed neural networks (PINNs) which jointly uses a Multilayer Perceptron (MLP) and Runge-Kutta method (RK) for solution estimation. The highlight of PINN is that the solution is regularized to satisfy the given PDE, by way of loss functions, with the help of automatic derivation made possible in modern software frameworks (PyTorch, TensorFlow).

I want to inform the world about AI and its impact on science. I am writing blogs which explains basic concepts of AI and related ground-breaking papers. I am also publishing papers within my research and related fields.

TITLE

Understanding spectral bias in PINNs

Spectral bias occurs when the model, mostly the architecture of the neural networks, used to solve the forward PDE problems tends to be biased towards low-frequency components of the function..

TITLE

Explaining the details of physics-informed neural networks

Physics-informed neural networks (PINNs) are a three-stage framework for solving continuous- and discrete-time PDEs. They combine a layered neural network to learn the solution, a physics-informed section to compute PDE residuals via automatic differentiation..