This repository highlights my contributions and independent learning to an advanced research project centered on solving differential equations in physics through the application of neural networks. The project leverages Physics-Informed Neural Networks (PINNs), employing both Multi-Layer Perceptrons (MLPs) and Kolmogorov-Arnold Networks (KANs) implemented in PyTorch, alongside conventional SciPy differential equation solving methods.
- More context + poster: Project page on my website
The primary focus of this research is to explore and elucidate the complex phenomenon of chaos in systems of driven damped pendulums. The objective is to rigorously compare the efficacy of MLPs, KANs, and traditional SciPy solvers in modeling these chaotic systems under various input conditions, thereby determining the most robust approach for accurately solving these equations.
Through this project, I have:
- Gained in-depth knowledge of chaotic systems and their behavior in physics.
- Enhanced my proficiency in PyTorch and SciPy for scientific computing.
notebooks/: Jupyter notebooks for experiments and analysis.notebooks/final/: “Best” / cleaned notebooks to read first.notebooks/experiments/: Parameter sweeps and work-in-progress runs.notebooks/legacy/: Older notebook copies kept for reference.
runs/: Saved artifacts from runs (CSV results, loss curves, plots, parameters).