I've trained a network on 10 nuclear configurations using n2p2. I input the data in the usual format with units of bohrs (length), hartrees (energy) and hartrees per bohr (force). The network trains successfully and when I use nnp-predict the energy of a different snapshot is in (close) agreement with the DFT prediction that I did as a benchmark. When I then copy the weights, scaling file and input.nn file across to cp2k to perform single point and molecular dynamic simulations I encounter problems. When I give cp2k the coordinates in Bohrs and cell parameters in bohr, the energy estimated is grossly incorrect ( -50 hartrees vs -2277 hartress approx). Does anyone know why this may be the case?
Please find attached:
input.nn - parameters for the neural network
input.data - data input file for my system in bohr, hartree and hartree/bohr for (length, energy and force)
scaling.data - the symmetry function file
energy.out - the output from nnp-predict
start.inp - the input file for cp2k using the weights and network parameters/functions for energy prediction
hybrid.out - the output file for cp2k showing the wrong energy prediction
files_for_question.zip
kjaj98
commented
1 year ago
I've trained a network on 10 nuclear configurations using n2p2. I input the data in the usual format with units of bohrs (length), hartrees (energy) and hartrees per bohr (force). The network trains successfully and when I use nnp-predict the energy of a different snapshot is in (close) agreement with the DFT prediction that I did as a benchmark. When I then copy the weights, scaling file and input.nn file across to cp2k to perform single point and molecular dynamic simulations I encounter problems. When I give cp2k the coordinates in Bohrs and cell parameters in bohr, the energy estimated is grossly incorrect ( -50 hartrees vs -2277 hartress approx). Does anyone know why this may be the case?
Please find attached: input.nn - parameters for the neural network input.data - data input file for my system in bohr, hartree and hartree/bohr for (length, energy and force) scaling.data - the symmetry function file energy.out - the output from nnp-predict start.inp - the input file for cp2k using the weights and network parameters/functions for energy prediction hybrid.out - the output file for cp2k showing the wrong energy prediction files_for_question.zip