K. T. Schütt, M. Gastegger, A. Tkatchenko, K.-R. Müller & R. J. Maurer (2019)
This paper had received a lot of attention, so I had to see what the fuzz was about. The method (SchNOrb) uses a deep neural network to create a Hamiltonian matrix from 3D coordinates, which can then be diagonalised to yield orbital energies and orbitals. SchNOrb is trained to reproduce Hamiltonian and overlap matrix elements, total energies (which are computed as the sum of orbital energies in the ML model), and gradients all taken from AIMD trajectories.
Highlighted by Jan Jensen
Figure 2 from the paper. © The Authors 2019. Reproduced under the CC-BY license.
This paper had received a lot of attention, so I had to see what the fuzz was about. The method (SchNOrb) uses a deep neural network to create a Hamiltonian matrix from 3D coordinates, which can then be diagonalised to yield orbital energies and orbitals. SchNOrb is trained to reproduce Hamiltonian and overlap matrix elements, total energies (which are computed as the sum of orbital energies in the ML model), and gradients all taken from AIMD trajectories.
So it's a bit like the ANI-1 method except that you also get orbitals, which can be used to compute other properties without additional parameterisation. One crucial difference though is that, as far as I can tell, SchNOrb is parameterised for each molecule separately.
The model uses about 93 million parameters for a >100 AO Hamiltonian and requires 25,000 structures and 80 GPU hours to train. Once trained it can predict an energy with meV accuracy in about 50 ms on a GPU.
The software is "available upon request".
The reviews are made available, and well worth reading.
This work is licensed under a Creative Commons Attribution 4.0 International License.