Skip to content
Technical University of Munich

Our research group software is available publicly on gitlab:
https://gitlab.com/fd-research

swimnetworks

With SWIM, neuron weights are sampled at high gradients of the target function.

Python package to quickly train feed-forward neural networks for supervised learning. This was developed in collaboration with Erik Bolager, Iryna Burak, Chinmay Datar, and Qing Sun. There is a paper associated with it: NeurIPS’23 and arXiv pre-print.

swim-rnn

The Lorenz attractor can be modelled accurately with SWIM-RNN.

Python package to quickly train recurrent neural networks without gradient descent. This was developed in collaboration with many people, see our pre-print associated with it: arXiv.

swim-pde

Solving the Helmholz equation on a complicated domain (a bunny).

Python package to solve partial differential equations with neural networks. This was developed in collaboration with many people, see our pre-print associated with it: arXiv.

swim-hgn

Network architecture of our Hamiltonian graph network.

Quickly train graph neural networks without gradient descent to learn Hamiltonian dynamics. See our pre-print associated with it on arXiv.

datafold

Eigenfunctions of the Laplace operator computed with diffusion maps.

Software for efficient manifold learning and time series data processing. This was developed in collaboration with Daniel Lehmberg. There is a paper associated with it: https://doi.org/10.21105/joss.02283.

NOMAD laboratory

Logo of NOMAD.

Developed in the FAIRmat project, enables scalable and FAIR sharing and use of research data.

Learning stochastic differential equations (SDEs)

Trajectories of an SIR model, learned with a neural network.

Using neural networks to learn stochastic differential equations from time series data. See the arXiv paper and the published version in CHAOS for a more detailed description.

To top