K = 2 # body order. K = 1 -> B1, K = 2 -> B2, K = 3 -> B3
descriptor_settings = [n_species, K, n_max, l_max]
desc = Bk(..., descriptor_settings)
The lammps files are changed correspondingly.
Another feature: add atom_indices to sparse_gp.py/update_db and sparse_gp.cpp/add_training_structure to allow adding a subset of force labels, such that the Kuf matrix is smaller in the "f" dimension
Notes:
To add descriptors with higher body order, just need to add coefficients (l1, l2, l3 ...; m1, m2, m3, ...) to coeffs.cpp and add indices (n1 l1 m1, n2 l2 m2, ... nk lk mk) in indices.cpp.
The original b1/2/3.cpp are kept for unit test, but we could delete them and might find better way of doing this.
[x] Finish unit test for lammps
[x] Change compute command for variance to Bk descriptors
Todo
The lammps code can be further optimized:
The trick as #12 is not implemented
The search of neighbor list and the computation of single bonds are duplicated for each kernel. But we can do it only once by using the maximal cutoff, n_max and l_max
This pull request wraps up B1, B2, B3 descriptors into a Bk descriptor module. And should be merged after #12
is equivalent to current setting
The lammps files are changed correspondingly.
atom_indicestosparse_gp.py/update_dbandsparse_gp.cpp/add_training_structureto allow adding a subset of force labels, such that the Kuf matrix is smaller in the "f" dimensionNotes:
To add descriptors with higher body order, just need to add coefficients (l1, l2, l3 ...; m1, m2, m3, ...) to
coeffs.cppand add indices (n1 l1 m1, n2 l2 m2, ... nk lk mk) inindices.cpp.The original
b1/2/3.cppare kept for unit test, but we could delete them and might find better way of doing this.computecommand for variance to Bk descriptorsTodo
The lammps code can be further optimized: