Author Topic: MTP——“No new candidates found in active learning MD”  (Read 1397 times)

0 Members and 1 Guest are viewing this topic.

Offline fangbiaolee

  • New QuantumATK user
  • *
  • Posts: 4
  • Country: cn
  • Reputation: 0
    • View Profile
Dear ATK-Staff/developer,

I used the AIMD  trajectory as the initial training set to generate the MTP potential, although I lowered the basis set of the MTP many times, the same problem still occurs--"  No new candidates found in active learning MD", is there any way to solve such a problem, or my initial training set has question?

part of the code:

trajectory=nlread('800K_0GPa_2ps.hdf5',MDTrajectory)[-1]
training_set = TrainingSet(trajectory,
                        sample_size=1000,
                        calculator=calculator)

fitting_parameters = MomentTensorPotentialFittingParameters(
    basis_size=200,
    inner_cutoff_radii=0.9 * Angstrom,
    tapering_cutoff_radii=1.1 * Angstrom,
    outer_cutoff_radii= 6.0 * Angstrom,
    constant_terms=constant_terms,
    mtp_filename='mtp_potential.mtp',
    forces_cap= 100.0 * eV / Angstrom,
    non_linear_coefficients_parameters=non_linear_coefficients_parameters,
)

 WARNING: The condition number of the optimized training matrix is            |
|          5.09e+13.This could lead to accuracy problems when calculating the  |
|          extrapolation grade.Consider reducing the number of MTP basis       |
|          functions or adding more diverse training configurations until the  |
|          condition number is lower.                                          |


Offline Julian Schneider

  • QuantumATK Staff
  • QuantumATK Guru
  • *****
  • Posts: 163
  • Country: dk
  • Reputation: 25
    • View Profile
Active learning can be a bit sensitive the size of the MTP basis, especially  if you have only few configurations in the initial training data or all configurations in the training data are relatively similar, which is often the case when you use a realtively short MD trajectory as initial training data.
In that case the given MTP basis can "overdescribe" the training data and the active learning algorithm becomes numerically unstable and cannot reliably select new candidate configurations. This is also what the warning says.

The first thing to try in this case, is to use a lower MTP basis, e.g. 50-100, until the warning disappears.
Once you have collected more training data, you can increase the number of basis functions again.

Alternatively, one can try to sample more different configurations in the initial training data, by running longer MD simulations, or using different initial confgurations for the MD.