Lastly, with the ability to analyze systems of these sizes, we can establish empirical convergence rates/scaling laws for growing system sizes. Consistent across two kinds of structures, we find a power law describing convergence.

Crucially, our method is not limited to biochemical compounds but also extends to computationally challenging organometallic compounds with up to 180 electrons where we - like the previous results - match or surpass the gold-standard CCSD(T)/CBS.

While, for a short cutoff of 5 Bohr radii, our approach matches the 'gold standard' CCSD(T)/CBS in non-covalent binding energies, an even shorter cutoff of 3 yields closer-to-experimental results in well-bonded biochemical compounds.

In the real world, this leads to speedups of up to 10x when scaling to a large number of electrons, as we do in this work. Thanks to VMC's embarrassingly parallel nature, this will enable the study of larger and more challenging systems as hardware progresses.

In this work, we set out to change this and improve the theoretical complexity of both operations by O(N).
We enable efficient wave function updates and efficient Laplacian computations by range-limiting electron-electron interactions within the neural network embedding.

Neural network VMC promises accuracy at scale but has been plagued with prohibitive costs mainly due to two reasons:
1) Densely connected neural wave functions cannot efficiently update if a few electrons' positions are changed.
2) Kinetic energy computations are expensive.

Accurate solutions to the electronic Schrödinger equation are a root problem in studying drugs and materials. As the number of electrons increases, the number of accurate methods grows thin.

Excited to present our work on Neural Pfaffians at #NeurIPS.

🗣️ Oral: Friday 3:30pm, East Ballroom A, B
📊 Post: Friday 4:30pm - 7:30pm, East Exhibit Hall A-C #3600
📝 Paper: openreview.net/forum?id=HRk...

Happy to chat!