EA Assignment 3 - PINNs and Neural ODEs

Full Name: Twamaa Makame
Assignment Title: EA Assignment 3
Roll Number: Z6003

Description

This repository contains solutions for EA Assignment 3 which focuses on:

1. Estimating cardiac activation times on a 2D domain using sparse data.
2. Comparing traditional feedforward neural networks with Physics-Informed Neural Networks (PINNs).
3. Using Neural Ordinary Differential Equations (Neural ODEs) for binary classification on a 2D dataset.

The code is organized into sections that generate synthetic data, define conduction velocity, train data-only and PINN models, visualize results, and build Neural ODE classifiers.

Repository Structure

/EA_Assignment3/
│
├── README.md <- This file
├── requirements.txt <- Dependencies for the project
│
├── /src/ <- Python scripts and notebooks
│ ├── pinn_model.py
│ ├── ode_model.py
│ ├── data_generation.py
│ └── visualization.py
│
└── /test/ <- Optional test cases and outputs
└── sample_inputs_outputs/

How to Run

1. Clone the repository: git clone[(https://github.com/twamaa/PINNs-and-Neural-ODE)]

2. Install dependencies: pip install -r requirements.txt

3. Open the Jupyter notebooks in the /src/ folder or run scripts directly for:

   • Generating synthetic data
   • Training and evaluating both models
   • Producing visualizations

Dependencies

See requirements.txt. Key packages include:

• numpy
• matplotlib
• torch
• torchdiffeq
• scikit-learn
• pyDOE

Known Issues

• The PINN model occasionally produces high RMSE if the sparse points do not sample critical wavefront boundaries.
• Neural ODE training can be slower than standard networks.

Assumptions

• Activation function and velocity maps are synthetic and idealized.
• Moons dataset used for Neural ODE comparison is noise-injected but simplified.