run_model.py: This can be used to apply the energy+force computation. It defaults to using the 'input.xyz' molecule. This will compute the energies and forces on that molecule using the trained NN. This can then be fed to a molecular dynamics program to adjust coordinates until the system converges.

422 test molecules were held out from training and used in analysis. The respective geometries are defined as:

422_mmff.xyz: a very fast optimization that might be a typical starting point for optimization (the default 'input.xyz' is the first molecule in this set as an example) 422_xd.xyz: an expensive optmimization (wB97XD) which was the target for the NN training. 422_nn.xyz: the trained neural-network optimization that converges quite close to the geometries of 422_xd.xyz.

The idea is to take a low level theory (eg mmff) to get a rough optimized geometry, and then progressively pass it through the NN using the method of the sample program. When it converges, the end geometry should approximate that of a much higher level of theory at a computation cost comparable to the level theory.