This is designed to interpolate chemical potential energy surfaces in the region around typical reaction paths.
Training data are hessians from quantum chemistry program outputs. A trained network takes as input, the x,y,z coords of each atom in the molecule.
180630: Works fine for energies, accuracy isn't yet good enough for gradients, once the keras team implements an L-BFGS optimizer things might look better (or I'll do it myself).
The readin.f90 file MUST be compiled on your machine before executing neuralnetwork_sean_conv.py using: f2py --fcompiler=gnu95 --f90flags=-fopenmp --opt=-O2 -c -m readin readin.f90 -llapack -lblas -lgomp