Deep learning for universal linear embeddings of nonlinear dynamics “Koopman operator”

Project for the Advance Machine Learning 2023 course

• Authors: Alessandro Mollicone, Giuliano Verrando
• Professor: Stefano Giagu
• Assistant: Andrea Ciardello
• University: "Sapienza Universita' di Roma"
• Academic Year: 2022-2023

---

---

The purpose of the project was to try to replicate the results obtained in this PAPER, in which the authors focused on developing DNN representations of Koopman eigenfunctions that remain interpretable and parsimonious, even for high-dimensional and strongly nonlinear systems. All data and methodologies can be reconstructed using the code available at GIT.

The authors make available all the data providing some matlab scripts to generate them. Alternatively, is possible to download them from here.

N.B. our Network notebook expects to find the data in a directory called data in the same location of the Network itself

The systems analized by the authors are:

1. A system described by a discrete spectrum
2. A fluid flowing in a box
3. A fluid flowing on an attractor
4. A non-linear pendulum

These are all examples of non-linear dynamics.