Transferable atomic orbitals

This release contains all code required to reproduce results of our papers „Towards a transferable fermionic neural wavefunction for molecules“ and “Variational Monte Carlo on a Budget — Fine-tuning pre-trained Neural Wavefunctions“.

https://doi.org/10.1038/s41467-023-44216-9 (Nature Communications)
https://openreview.net/forum?id=FBNyccPfAu

Among other features it supports:

• Transferable atomic orbitals (TAO) to optimize a single neural network across molecules and geometrical conformations.
• Normal mode distortion for geometrical conformations to sample continously new geometrical conformations during shared optimization.
• A variation of PhisNet to generate orbital features for the TAO ansatz including a training framework.
• A dataset of molecules up to a size of four heavy atoms (Oxygen, Carbon, Nitrogen) for the pre-training of a TAO wavefunction.

While still contained in the codebase, some older features not relevant for the paper (such as weight-sharing or PauliNet-like architectures) have not been tested and might be broken in this release.

arxiv_2205.09438v2

This release contains all code to reproduce our arxiv-paper
Gold-standard solutions to the Schrödinger equation using deep learning: How much physics do we need?
https://arxiv.org/abs/2205.09438v2

Among other things it supports:

• PauliNet-like-architectures, FermiNet-like-architectures, as well as our own improved architecture detailed in the paper
• Optimization using KFAC
• Local coordinate systems
• Physics-inspired initializiation of envelopes

While still contained in the codebase, some older features not relevant for the paper (such as weight-sharing) have not been tested and might be broken in this release.

arxiv_2105.08351v2

This release contains all code required to reproduce results of our paper
Solving the electronic Schrödinger equation for multiple nuclear geometries with weight-sharing deep neural networks

https://doi.org/10.1038/s43588-022-00228-x (Nature Computational Science)
https://arxiv.org/abs/2105.08351 (arxiv)

Among other features it supports:

• Weight-sharing optimization across multiple molecules
• Re-use of weights for different geometries/molecules
• PauliNet-like architectures and our improved version DeepErwin

For an improved architecture, with better accuracy please refer to our more recent releases, such as
https://github.com/mdsunivie/deeperwin/releases/tag/arxiv_2205.09438v2