Question on the perturb function

[yunjian-hash]

Dear developers,

When the cell is perturbed, will the atomic coordinates change accordingly? In addition, why perturb() function must require a positional argument: 'atom_pert_distance', if I only want the cell changes and the coordinates of atoms change accordingly, should I set the atom_pert_distance equals to 0?

Thanks!

[HuangJiameng]

Atoms will move about atom_pert_distance in random direction. And it can be set as 0. Source code is here, FYI.

dpdata/dpdata/system.py

Lines 774 to 826 in 2ac2f48

	def perturb(self,
	pert_num,
	cell_pert_fraction,
	atom_pert_distance,
	atom_pert_style='normal'):
	"""
	Perturb each frame in the system randomly.
	The cell will be deformed randomly, and atoms will be displaced by a random distance in random direction.
	Parameters
	----------
	pert_num : int
	Each frame in the system will make `pert_num` copies,
	and all the copies will be perturbed.
	That means the system to be returned will contain `pert_num` * frame_num of the input system.
	cell_pert_fraction : float
	A fraction determines how much (relatively) will cell deform.
	The cell of each frame is deformed by a symmetric matrix perturbed from identity.
	The perturbation to the diagonal part is subject to a uniform distribution in [-cell_pert_fraction, cell_pert_fraction),
	and the perturbation to the off-diagonal part is subject to a uniform distribution in [-0.5*cell_pert_fraction, 0.5*cell_pert_fraction).
	atom_pert_distance: float
	unit: Angstrom. A distance determines how far atoms will move.
	Atoms will move about `atom_pert_distance` in random direction.
	The distribution of the distance atoms move is determined by atom_pert_style
	atom_pert_style : str
	Determines the distribution of the distance atoms move is subject to.
	Avaliable options are
	- `'normal'`: the `distance` will be object to `chi-square distribution with 3 degrees of freedom` after normalization.
	The mean value of the distance is `atom_pert_fraction*side_length`
	- `'uniform'`: will generate uniformly random points in a 3D-balls with radius as `atom_pert_distance`.
	These points are treated as vector used by atoms to move.
	Obviously, the max length of the distance atoms move is `atom_pert_distance`.
	- `'const'`: The distance atoms move will be a constant `atom_pert_distance`.
	Returns
	-------
	perturbed_system : System
	The perturbed_system. It contains `pert_num` * frame_num of the input system frames.
	"""
	perturbed_system = System()
	nframes = self.get_nframes()
	for ii in range(nframes):
	for jj in range(pert_num):
	tmp_system = self[ii].copy()
	cell_perturb_matrix = get_cell_perturb_matrix(cell_pert_fraction)
	tmp_system.data['cells'][0] = np.matmul(tmp_system.data['cells'][0],cell_perturb_matrix)
	tmp_system.data['coords'][0] = np.matmul(tmp_system.data['coords'][0],cell_perturb_matrix)
	for kk in range(len(tmp_system.data['coords'][0])):
	atom_perturb_vector = get_atom_perturb_vector(atom_pert_distance, atom_pert_style)
	tmp_system.data['coords'][0][kk] += atom_perturb_vector
	tmp_system.rot_lower_triangular()
	perturbed_system.append(tmp_system)
	return perturbed_system

[yunjian-hash]

When the cell is perturbed, will the atomic coordinates change accordingly?

Thanks! How about the first questrion 'when the cell is perturbed, will the atomic coordinates change accordingly?', the code does not tell about that.

[HuangJiameng]

I think that it does tell.

dpdata/dpdata/system.py

Lines 819 to 820 in 2ac2f48

	tmp_system.data['cells'][0] = np.matmul(tmp_system.data['cells'][0],cell_perturb_matrix)
	tmp_system.data['coords'][0] = np.matmul(tmp_system.data['coords'][0],cell_perturb_matrix)

[yunjian-hash]

Thanks! When I convert the system after pertubed, here comes an error. are the perturbed systems cannot be append to the multisystem?

from dpdata import System, LabeledSystem, MultiSystems
import numpy as np

####### 读入轨迹 #######

ls3=LabeledSystem('./Walker_2', fmt='cp2k/aimd_output')

####### 验证集结构数 #######

a3 = np.random.choice(len(ls3),size=100,replace=False)

####### 训练集结构数 #######

w3 = list(set(range(1,len(ls3)))-set(a3))[10:-1:6]
rs3=ls3.sub_system(w3)

####### 微扰训练集 #######
perturbed_system_1 = rs3.perturb(pert_num=1, cell_pert_fraction=0.1, atom_pert_distance=0, atom_pert_style='normal')
perturbed_system_2 = rs3.perturb(pert_num=1, cell_pert_fraction=0.2, atom_pert_distance=0, atom_pert_style='normal')
perturbed_system_3 = rs3.perturb(pert_num=1, cell_pert_fraction=0.3, atom_pert_distance=0, atom_pert_style='normal')

####### 训练集 #######
ms= MultiSystems()
ms.append(rs3)
ms.append(perturbed_system_1)
ms.append(perturbed_system_2)
ms.append(perturbed_system_3)

ms.to_deepmd_raw('deepmd')
ms.to_deepmd_npy('deepmd',set_size=5000)

print('There are %d frames for training' % len(ls3.sub_system(w3)))
print('There are 3* %d frames for training with perturbed system' % len(perturbed_system_1))

[LiangWenshuo1118]

I guess this is because perturbed_systems only have coords and cells. Energies, forces and virials information should also be contained.

[wanghan-iapcm]

referred to #381