motep train

This command trains the potential starting from potentials.initial based on configurations.training. The trained potential is written in potentials.final.

Usage

motep train motep.train.toml

or

mpirun -np 4 motep train motep.train.toml

motep.train.toml

[common]
seed = 42  # random seed for initializing MTP parameters etc.

[configurations]
training = 'training.cfg'

[potentials]
initial = 'initial.mtp'
final = 'final.mtp'

[loss]  # setting for the loss function
energy_weight = 1.0
forces_weight = 0.01
stress_weight = 0.001

# optimization steps

# style 1: simple
# steps = ['Nelder-Mead', 'BFGS']

# style 2: sophisticated
# "optimized" specifies which parameters are optimized at the step.
[[steps]]
method = 'Nelder-Mead'
optimized = ['species_coeffs', 'radial_coeffs', 'moment_coeffs']
[steps.kwargs]
tol = 1e-7
[steps.kwargs.options]
maxiter = 1000

[[steps]]
method = 'BFGS'
optimized = ['species_coeffs', 'radial_coeffs', 'moment_coeffs']

If some of the following parameters are already given in initial.mtp, they are treated as the initial guess, which may or may not be optimized depending on the above setting.

• scaling (not recommended to optimized)

• radial_coeffs

• moment_coeffs

• species_coeffs