Physics Informed Neural Network
Project Description
This project explores the development and application of Physics-Informed Neural Networks (PINNs) to solve complex differential equations governing physical systems. By integrating known physical laws (e.g., conservation laws, PDEs) directly into the neural network’s loss function, PINNs leverage both data-driven learning and physics-based constraints, enabling accurate predictions even with sparse or noisy data. The study focuses on optimizing PINN architectures, training strategies, and hybrid approaches to improve convergence and generalization across domains such as fluid dynamics, heat transfer, and structural mechanics. Through numerical experiments and benchmark comparisons, the project aims to demonstrate PINNs’ effectiveness in solving forward and inverse problems while reducing reliance on high-fidelity simulations. Key challenges, such as handling multi-scale phenomena and improving training efficiency, will also be investigated to advance the adoption of PINNs in scientific and engineering applications.
Supervisor
LAU, Kin Nang
Quota
3
Course type
UROP3100
UROP3200
UROP4100
Applicant's Roles
The student will conduct research and implement the solution of PINN to find the strong and weak solution of one type of PDE system. The domains may be Quantum Mechanics (Schrodinger equations), String theory (Cosmic String equations) or Fluid Mechanics (Naiver Strokes equations). Several target foci may include (a) Finding weak solution when the solution space of PDE involves non-differentiable functions or even discontinuous functions.
(b) Finding symbolic solution using KAN and symbolic regression
Applicant's Learning Objectives
(a) understanding PDE systems
(b) understanding how AI is applied to find solutions of PDE systems
(c) understanding how to train a NN using auto-differentiation
(d) understanding how to do symbolic regression
Complexity of the project
Challenging