Source code for pyxtal.lego.SO3

from __future__ import division
import numpy as np
from scipy.special import sph_harm, spherical_in
from ase import Atoms
from ase.neighborlist import NeighborList


[docs] class SO3: ''' A class to generate the SO3 power spectrum components based off of the Gaussian atomic neighbor density function defined in "On Representing Atomic Environments". args: nmax: int, degree of radial expansion lmax: int, degree of spherical harmonic expansion rcut: float, cutoff radius for neighbor calculation alpha: float, gaussian width parameter weight_on: bool, if True, the neighbors with different type will be counted as negative ''' def __init__(self, nmax=3, lmax=3, rcut=3.5, alpha=2.0, weight_on=False): # populate attributes self.nmax = nmax self.lmax = lmax self.rcut = rcut self.alpha = alpha self._type = "SO3" self.cutoff_function = 'cosine' self.weight_on = weight_on self.ncoefs = self.nmax*(self.nmax+1)//2*(self.lmax+1) self.tril_indices = np.tril_indices(self.nmax, k=0) self.ls = np.arange(self.lmax+1) self.norm = np.sqrt(2*np.sqrt(2)*np.pi/np.sqrt(2*self.ls+1)) self.keys = ['keys', '_nmax', '_lmax', '_rcut', '_alpha', '_cutoff_function', 'weight_on', 'ncoefs', 'ls', 'norm', 'tril_indices', 'args'] self.args = (self.nmax, self.lmax, self.rcut, self.alpha, self._cutoff_function) def __str__(self): s = "SO3 descriptor with Cutoff: {:6.3f}".format(self.rcut) s += " lmax: {:d}, nmax: {:d}, alpha: {:.3f}\n".format(self.lmax, self.nmax, self.alpha) return s def __repr__(self): return str(self)
[docs] def load_from_dict(self, dict0): self.nmax = dict0["nmax"] self.lmax = dict0["lmax"] self.rcut = dict0["rcut"] self.alpha = dict0["alpha"] self.derivative = dict0["derivative"]
[docs] def save_dict(self): """ save the model as a dictionary in json """ dict = {"nmax": self.nmax, "lmax": self.lmax, "rcut": self.rcut, "alpha": self.alpha, "derivative": self.derivative, "_type": "SO3", } return dict
@property def nmax(self): return self._nmax @nmax.setter def nmax(self, nmax): if isinstance(nmax, int) is True: if nmax < 1: raise ValueError('nmax must be greater than or equal to 1') if nmax > 11: raise ValueError('nmax > 11 yields complex eigenvalues') self._nmax = nmax else: raise ValueError('nmax must be an integer') @property def lmax(self): return self._lmax @lmax.setter def lmax(self, lmax): if isinstance(lmax, int) is True: if lmax < 0: raise ValueError('lmax must be greater than or equal to zero') elif lmax > 32: raise NotImplementedError('support a maxmimum l=32 for spherical harmonics') self._lmax = lmax else: raise ValueError('lmax must be an integer') @property def rcut(self): return self._rcut @rcut.setter def rcut(self, rcut): if isinstance(rcut, float) or isinstance(rcut, int): if rcut <= 0: raise ValueError('rcut must be greater than zero') self._rcut = rcut else: raise ValueError('rcut must be a float') @property def alpha(self): return self._alpha @alpha.setter def alpha(self, alpha): if isinstance(alpha, float) or isinstance(alpha, int): if alpha <= 0: raise ValueError('alpha must be greater than zero') self._alpha = alpha else: raise ValueError('alpha must be a float') @property def cutoff_function(self): return self._cutoff_function @cutoff_function.setter def cutoff_function(self, cutoff_function): self._cutoff_function = Cosine
[docs] def clear_memory(self): ''' Clears all non essential attributes for the calculator ''' attrs = list(vars(self).keys()) for attr in attrs: if attr not in self.keys: delattr(self, attr) return
[docs] def init_atoms(self, atoms, atom_ids=None): """ initilize atoms related attributes """ self._atoms = atoms self.natoms = len(atoms) self.build_neighbor_list(atom_ids)
[docs] def compute_p(self, atoms, atom_ids=None, return_CN=False): """ Compute the powerspectrum function Args: atoms: ase atoms object atom_ids: optional list of atomic indices Returns: p array (N, M) """ if atom_ids is None: atom_ids = range(len(atoms)) self.init_atoms(atoms, atom_ids) plist = np.zeros((len(atom_ids), self.ncoefs), dtype=np.float64) dist = np.zeros((len(atom_ids), 2)) if len(self.neighborlist) > 0: cs = compute_cs(self.neighborlist, *self.args) cs *= self.atomic_weights[:, np.newaxis, np.newaxis, np.newaxis] cs = np.einsum('inlm,l->inlm', cs, self.norm) # Get r_ij and compute C*np.conj(C) for _i, i in enumerate(atom_ids): centers = self.neighbor_indices[:, 0] == i CN = len(self.neighborlist[centers]) if CN > 0: ctot = cs[centers].sum(axis=0) P = np.einsum('ijk,ljk->ilj', ctot, np.conj(ctot)).real plist[_i] = P[self.tril_indices].flatten() dist[_i, 0] += np.linalg.norm(self.neighborlist[centers], axis=1).max() dist[_i, 1] += CN if return_CN: return plist, dist else: return plist
[docs] def compute_dpdr(self, atoms, atom_ids=None): """ Compute the powerspectrum function Args: atoms: ase atoms object atom_ids: optional list of atomic indices Returns: dpdr array (N, N, M, 3) and p array (N, M) """ if atom_ids is None: atom_ids = range(len(atoms)) self.init_atoms(atoms, atom_ids) p_list = np.zeros((len(atom_ids), self.ncoefs), dtype=np.float64) dp_list = np.zeros((len(atom_ids), self.natoms, self.ncoefs, 3), dtype=np.float64) if len(self.neighborlist) > 0: # get expansion coefficients and derivatives cs, dcs = compute_dcs(self.neighborlist, *self.args) # weight cs and dcs cs *= self.atomic_weights[:, np.newaxis, np.newaxis, np.newaxis] dcs *= self.atomic_weights[:, np.newaxis, np.newaxis, np.newaxis, np.newaxis] cs = np.einsum('inlm,l->inlm', cs, self.norm) dcs = np.einsum('inlmj,l->inlmj', dcs, self.norm) #print('cs, dcs', self.neighbor_indices, cs.shape, dcs.shape) # cs: (N_ij, n, l, m) => P (N_i, N_des) # dcs: (N_ij, n, l, m, 3) => dP (N_i, N_j, N_des, 3) # (n, l, m) needs to be merged to 1 dimension for _i, i in enumerate(atom_ids): # find atoms for which i is the center centers = self.neighbor_indices[:, 0] == i if len(self.neighbor_indices[centers]) > 0: # total up the c array for the center atom ctot = cs[centers].sum(axis=0) #(n, l, m) # power spectrum P = c*c_conj # eq_3 (n, n', l) eliminate m P = np.einsum('ijk, ljk->ilj', ctot, np.conj(ctot)).real p_list[_i] += P[self.tril_indices].flatten() # gradient of P for each neighbor, eq_26 # (N_ijs, n, n', l, 3) # dc * c_conj + c * dc_conj dP = np.einsum('wijkn,ljk->wiljn', dcs[centers], np.conj(ctot)) dP += np.conj(np.transpose(dP, axes=[0, 2, 1, 3, 4])) dP = dP.real #print("shape of P/dP", P.shape, dP.shape)#; import sys; sys.exit() # QZ: to check ijs = self.neighbor_indices[centers] for _id, j in enumerate(ijs[:, 1]): tmp = dP[_id][self.tril_indices].flatten().reshape(self.ncoefs, 3) dp_list[_i, j, :, :] += tmp dp_list[_i, i, :, :] -= tmp return dp_list, p_list
[docs] def compute_dpdr_5d(self, atoms, atom_ids=None): """ Compute the powerspectrum function with respect to supercell Args: atoms: ase atoms object Returns: dpdr array (N, N, M, 3, 27) and p array (N, M) """ if atom_ids is None: atom_ids = range(len(atoms)) self.init_atoms(atoms, atom_ids) p_list = np.zeros((len(atom_ids), self.ncoefs), dtype=np.float64) dp_list = np.zeros((len(atom_ids), self.natoms, self.ncoefs, 3, 27), dtype=np.float64) if len(self.neighborlist) > 0: # get expansion coefficients and derivatives cs, dcs = compute_dcs(self.neighborlist, *self.args) # weight cs and dcs cs *= self.atomic_weights[:, np.newaxis, np.newaxis, np.newaxis] dcs *= self.atomic_weights[:, np.newaxis, np.newaxis, np.newaxis, np.newaxis] cs = np.einsum('inlm,l->inlm', cs, self.norm) dcs = np.einsum('inlmj,l->inlmj', dcs, self.norm) #print('cs, dcs', self.neighbor_indices, cs.shape, dcs.shape) # cs: (N_ij, n, l, m) => P (N_i, N_des) # dcs: (N_ij, n, l, m, 3) => dP (N_i, N_j, N_des, 3) # (n, l, m) needs to be merged to 1 dimension neigh_ids = np.arange(len(self.neighbor_indices)) for _i, i in enumerate(atom_ids): # find atoms for which i is the center pair_ids = neigh_ids[self.neighbor_indices[:, 0] == i] if len(pair_ids) > 0: ctot = cs[pair_ids].sum(axis=0) #(n, l, m) dctot = dcs[pair_ids].sum(axis=0) # power spectrum P = c * c_conj P = np.einsum('ijk, ljk->ilj', ctot, np.conj(ctot)).real p_list[_i] += P[self.tril_indices].flatten() # loop over each pair for pair_id in pair_ids: (_, j, cell_id) = self.neighbor_indices[pair_id] # map from (x, y, z) to (0, 27) # (N_ijs, n, n', l, 3) # dp = dc * c_conj + c * dc_conj dP = np.einsum('ijkn, ljk->iljn', dcs[pair_id], np.conj(ctot)) dP += np.einsum('ijkn, ljk->iljn', np.conj(dcs[pair_id]), ctot) dP = dP.real[self.tril_indices].flatten().reshape(self.ncoefs, 3) #print(cs[pair_id].shape, dcs[pair_id].shape, dP.shape) dp_list[_i, j, :, :, cell_id] += dP dp_list[_i, i, :, :, 13] -= dP return dp_list, p_list
[docs] def calculate(self, atoms, atom_ids=None, derivative=False): ''' API for Calculating the SO(3) power spectrum components of the smoothened atomic neighbor density function Args: atoms: an ASE atoms object corresponding to the desired atomic arrangement derivative: bool, whether to calculate the gradient of not ''' p_list = None dp_list = None if derivative: dp_list, p_list = self.compute_dpdr(atoms, atom_ids) else: p_list = self.compute_p(atoms, atom_ids) x = {'x': p_list, 'dxdr': dp_list, 'elements': list(atoms.symbols)} self.clear_memory() return x
[docs] def build_neighbor_list(self, atom_ids=None): ''' Builds a neighborlist for the calculation of bispectrum components for a given ASE atoms object given in the calculate method. ''' atoms = self._atoms cell_matrix = atoms.get_cell() VECTORS = np.array([[x1, y1, z1] for x1 in range(-1, 2) for y1 in range(-1, 2) for z1 in range(-1, 2)]) neighbors = [] neighbor_indices = [] atomic_weights = [] #if True: #atom_ids is None: if atom_ids is None: atom_ids = range(len(atoms)) cutoffs = [self.rcut/2]*len(atoms) nl = NeighborList(cutoffs, self_interaction=False, bothways=True, skin=0.0) nl.update(atoms) for i in atom_ids: # get center atom position vector center_atom = atoms.positions[i] # get indices and cell offsets for each neighbor indices, offsets = nl.get_neighbors(i) #print(indices); import sys; sys.exit() for j, offset in zip(indices, offsets): (x, y, z) = offset cell_id = (x+1) * 9 + (y+1) * 3 + z + 1 pos = atoms.positions[j] + offset@cell_matrix - center_atom if np.sum(np.abs(pos)) < 1e-3: # to skip self if cell_id == 13: continue # to prevent division by zero else: pos += 1e-3 neighbors.append(pos) if self.weight_on and atoms[j].number != atoms[i].number: factor = -1 else: factor = 1 atomic_weights.append(factor*atoms[j].number) neighbor_indices.append([i, j, cell_id]) else: # A short cut version if we only compute the neighbors for a few atoms ref_pos = np.repeat(atoms.positions[:, :, np.newaxis], 27, axis=2) cell_shifts = np.dot(VECTORS, cell_matrix) ref_pos += cell_shifts.T[np.newaxis, :, :] ref_ids = np.arange(len(atoms)) for i in atom_ids: center_atom = atoms.positions[i] # Compute distances relative to the center_atom across all cells dists = np.linalg.norm(ref_pos - center_atom[:, np.newaxis], axis=1) # Create mask for distances that are within the cutoff and greater than 1e-2 mask = (dists > 1e-2) & (dists < self.rcut) # Use the mask to find neighbors in all cells at once valid_atoms, valid_cells = np.where(mask) # Append the neighbors and weights for j, cell_id in zip(valid_atoms, valid_cells): neighbors.append(ref_pos[j, :, cell_id] - center_atom) factor = -1 if self.weight_on and atoms[j].number != atoms[i].number else 1 atomic_weights.append(factor * atoms[j].number) neighbor_indices.append([i, j, cell_id]) #for cell_id in range(27): # dists = np.linalg.norm(ref_pos[:, :, cell_id] - center_atom, axis=1) # mask = (dists > 1e-2) & (dists < self.rcut) # for j in ref_ids[mask]: # neighbors.append(ref_pos[j, :, cell_id] - center_atom) # factor = -1 if self.weight_on and atoms[j].number != atoms[i].number else 1 # atomic_weights.append(factor*atoms[j].number) # neighbor_indices.append([i, j, cell_id]) self.neighborlist = np.array(neighbors, dtype=np.float64) self.atomic_weights = np.array(atomic_weights, dtype=np.int64) self.neighbor_indices = np.array(neighbor_indices, dtype=np.int64)
[docs] def Cosine(Rij, Rc, derivative=False): # Rij is the norm if not derivative: result = 0.5 * (np.cos(np.pi * Rij / Rc) + 1.) else: result = -0.5 * np.pi / Rc * np.sin(np.pi * Rij / Rc) return result
[docs] def W(nmax): arr = np.zeros((nmax,nmax), np.float64) for alpha in range(1, nmax+1, 1): temp1 = (2*alpha+5)*(2*alpha+6)*(2*alpha+7) for beta in range(1, alpha+1, 1): temp2 = (2*beta+5)*(2*beta+6)*(2*beta+7) arr[alpha-1, beta-1] = np.sqrt(temp1*temp2)/(5+alpha+beta)/(6+alpha+beta)/(7+alpha+beta) arr[beta-1, alpha-1] = arr[alpha-1, beta-1] sinv = np.linalg.inv(arr) eigvals, V = np.linalg.eig(sinv) sqrtD = np.diag(np.sqrt(eigvals)) arr[:,:] = np.dot(np.dot(V, sqrtD), np.linalg.inv(V)) return arr
[docs] def phi(r, alpha, rcut): ''' See g below ''' return (rcut-r)**(alpha+2)/np.sqrt(2*rcut**(2*alpha+7)/(2*alpha+5)/(2*alpha+6)/(2*alpha+7))
[docs] def g(r, n, nmax, rcut, w): Sum = 0.0 for alpha in range(1, nmax+1): Sum += w[n-1, alpha-1]*phi(r, alpha, rcut) return Sum
[docs] def GaussChebyshevQuadrature(nmax, lmax): NQuad = (nmax+lmax+1) * 10 # quad_array = np.zeros(NQuad, dtype=np.float64) for i in range(1, NQuad+1): # roots of Chebyshev polynomial of degree N x = np.cos((2*i-1)*np.pi/2/NQuad) quad_array[i-1] = x return quad_array, np.pi/NQuad
[docs] def compute_cs(pos, nmax, lmax, rcut, alpha, cutoff): """ Compute expansion coefficients for a system based on the input positions. This function calculates the expansion coefficients for a set of atomic positions using Gauss-Chebyshev quadrature, spherical Bessel functions, and spherical harmonics. It is typically used in models that require high-dimensional projections of atomic environments. Args: pos (numpy.ndarray): An array of atomic positions (N x 3) where N is the number of atoms. nmax (int): Maximum radial quantum number used in the expansion. lmax (int): Maximum angular momentum quantum number used in the expansion. rcut (float): Cutoff radius for interactions and the radial expansion. alpha (float): Gaussian decay factor applied to the radial functions. cutoff (callable): A function to compute cutoff values for the radial distances. Returns: numpy.ndarray: A 4D array of expansion coefficients with shape (N_neighbors, nmax, lmax+1, 2*lmax+1), where `N_neighbors` is the number of neighbor atoms, `nmax` is the number of radial terms, and `lmax` and `m` correspond to angular momentum quantum numbers. """ # 1. Init Overlap matrix, distances and quadrature points w = W(nmax) Ris = np.linalg.norm(pos, axis=1) # (N_neighbors) GCQuadrature, weight = GaussChebyshevQuadrature(nmax, lmax) # (Nquad) weight *= rcut / 2 Quadrature = rcut / 2 * (GCQuadrature + 1) # 2. Rdial G functions Gs = np.zeros((nmax, len(Quadrature)), dtype=np.float64) # (nmax, Nquad) for n in range(1, nmax + 1): Gs[n - 1, :] = g(Quadrature, n, nmax, rcut, w) Quad_Squared = Quadrature ** 2 Gs *= Quad_Squared * np.exp(-alpha * Quad_Squared) * np.sqrt(1 - GCQuadrature**2) * weight # 3. Bessel functions BesselArgs = 2 * alpha * np.outer(Ris, Quadrature) # (N_neighbors x Nquad) Bessels = np.zeros((len(Ris), len(Quadrature), lmax + 1), dtype=np.float64) # (N_neighbors x Nquad x lmax+1) for l in range(lmax + 1): Bessels[:, :, l] = spherical_in(l, BesselArgs) # 4. Integration over radial coordinates using Einstein summation integral_array = np.einsum('ij,kjl->kil', Gs, Bessels) # (N_neighbors x nmax x lmax+1) # 5. Gaussian factors for each atom and apply the cutoff function exparray = 4 * np.pi * np.exp(-alpha * Ris**2) # (N_neighbors) cutoff_array = cutoff(Ris, rcut) np.multiply(exparray, cutoff_array, out=exparray) # 6. Compute the spherical harmonics for each atom # From Cartesian coordinates to spherical coordinates thetas = pos[:, 2] / Ris[:] thetas = np.arccos(thetas) phis = np.arctan2(pos[:, 1], pos[:, 0]) msize = 2 * lmax + 1 ylms = np.zeros((len(Ris), lmax + 1, msize), dtype=np.complex128) #for l in range(lmax + 1): # for m in range(-l, l + 1): # midx = msize // 2 + m # ylms[:, l, midx] = sph_harm(m, l, phis, thetas) l_vals = np.repeat(np.arange(lmax + 1), [2 * l + 1 for l in range(lmax + 1)]) m_vals = np.concatenate([np.arange(-l, l + 1) for l in range(lmax + 1)]) midx_vals = msize // 2 + m_vals ylms[:, l_vals, midx_vals] = sph_harm(m_vals, l_vals, phis[:, np.newaxis], thetas[:, np.newaxis]) # 7. Multiply Ylm and Gaussian factors Y_mul_innerprod = np.einsum('ijk,ilj->iljk', ylms, integral_array) C = np.einsum('i,ijkl->ijkl', exparray, Y_mul_innerprod) return C
[docs] def compute_dcs(pos, nmax, lmax, rcut, alpha, cutoff): """ Compute exapnsion coefficients Args: pos: nmax (int): lmax (int): rcut (float): alpha (float): cutoff (callable): Returns: c(N_ij, nmax, lmax+1, 2lmax+1) dc(N_ij, nmax, lmax+1, 2lmax+1, 3) for each x,y,z """ # compute the overlap matrix w = W(nmax) # get the norm of the position vectors Ris = np.linalg.norm(pos, axis=1) # (Nneighbors) # get unit vectors upos = pos/Ris[:,np.newaxis] # initialize Gauss Chebyshev Quadrature GCQuadrature, weight = GaussChebyshevQuadrature(nmax,lmax) #(Nquad) weight *= rcut/2 # transform from (-1,1) to (0, rcut) Quadrature = rcut/2*(GCQuadrature+1) # compute the arguments for the bessel functions BesselArgs = 2*alpha*np.outer(Ris,Quadrature)#(Nneighbors x Nquad) # initalize the arrays for the bessel function values # and the G function values Bessels = np.zeros((len(Ris), len(Quadrature), lmax+1), dtype=np.float64) #(Nneighbors x Nquad x lmax+1) Gs = np.zeros((nmax, len(Quadrature)), dtype=np.float64) # (nmax, nquad) dBessels = np.zeros((len(Ris), len(Quadrature), lmax+1), dtype=np.float64) #(Nneighbors x Nquad x lmax+1) # compute the g values for n in range(1,nmax+1,1): Gs[n-1,:] = g(Quadrature, n, nmax, rcut,w)*weight # compute the bessel values for l in range(lmax+1): Bessels[:,:,l] = spherical_in(l, BesselArgs) dBessels[:,:,l] = spherical_in(l, BesselArgs, derivative=True) #(Nneighbors x Nquad x lmax+1) unit vector here gradBessels = np.einsum('ijk,in->ijkn',dBessels,upos) gradBessels *= 2*alpha # multiply with r for the integral gradBessels = np.einsum('ijkn,j->ijkn',gradBessels,Quadrature) # mutliply the terms in the integral separate from the Bessels Quad_Squared = Quadrature**2 Gs *= Quad_Squared * np.exp(-alpha*Quad_Squared) * np.sqrt(1-GCQuadrature**2) # perform the integration with the Bessels integral_array = np.einsum('ij,kjl->kil', Gs, Bessels) # (Nneighbors x nmax x lmax+1) grad_integral_array = np.einsum('ij,kjlm->kilm', Gs, gradBessels)# (Nneighbors x nmax x lmax+1, 3) # compute the gaussian for each atom exparray = 4*np.pi*np.exp(-alpha*Ris**2) # (Nneighbors) gradexparray = (-2*alpha*Ris*exparray)[:,np.newaxis]*upos cutoff_array = cutoff(Ris, rcut) grad_cutoff_array = np.einsum('i,in->in',cutoff(Ris, rcut, True), upos) # get the spherical coordinates of each atom thetas = np.arccos(pos[:,2]/Ris[:]) phis = np.arctan2(pos[:,1], pos[:,0]) # the size changes temporarily for the derivative # determine the size of the m axis Msize = 2*(lmax+1)+1 msize = 2*lmax + 1 # initialize an array for the spherical harmonics and gradients #(Nneighbors, l, m, *3*) ylms = np.zeros((len(Ris), lmax+1+1, Msize), dtype=np.complex128) gradylms = np.zeros((len(Ris), lmax+1, msize, 3), dtype=np.complex128) # compute the spherical harmonics for l in range(lmax+1+1): for m in range(-l,l+1,1): midx = Msize//2 + m ylms[:,l,midx] = sph_harm(m, l, phis, thetas) for l in range(1, lmax+1): for m in range(-l, l+1, 1): midx = msize//2 + m Midx = Msize//2 + m # get gradient with recpect to spherical covariant components xcov0 = -np.sqrt(((l+1)**2-m**2)/(2*l+1)/(2*l+3))*l*ylms[:,l+1,Midx]/Ris if abs(m) <= l-1: xcov0 += np.sqrt((l**2-m**2)/(2*l-1)/(2*l+1))*(l+1)*ylms[:,l-1,Midx]/Ris xcovpl1 = -np.sqrt((l+m+1)*(l+m+2)/2/(2*l+1)/(2*l+3))*l*ylms[:,l+1,Midx+1]/Ris if abs(m+1) <= l-1: xcovpl1 -= np.sqrt((l-m-1)*(l-m)/2/(2*l-1)/(2*l+1))*(l+1)*ylms[:,l-1,Midx+1]/Ris xcovm1 = -np.sqrt((l-m+1)*(l-m+2)/2/(2*l+1)/(2*l+3))*l*ylms[:,l+1,Midx-1]/Ris if abs(m-1) <= l-1: xcovm1 -= np.sqrt((l+m-1)*(l+m)/2/(2*l-1)/(2*l+1))*(l+1)*ylms[:,l-1,Midx-1]/Ris #transform the gradient to cartesian gradylms[:,l,midx,0] = 1/np.sqrt(2)*(xcovm1-xcovpl1) gradylms[:,l,midx,1] = 1j/np.sqrt(2)*(xcovm1+xcovpl1) gradylms[:,l,midx,2] = xcov0 # index ylms to get rid of extra terms for derivative ylms = ylms[:, 0:lmax+1, 1:1+2*lmax+1] # multiply the spherical harmonics and the radial inner product Y_mul_innerprod = np.einsum('ijk,ilj->iljk', ylms, integral_array) # multiply the gradient of the spherical harmonics with the radial inner get_radial_inner_product dY_mul_innerprod = np.einsum('ijkn,ilj->iljkn', gradylms, integral_array) # multiply the spherical harmonics with the gradient of the radial inner get_radial_inner_product Y_mul_dinnerprod = np.einsum('ijk,iljn->iljkn', ylms, grad_integral_array) # multiply the gaussians into the expression with 4pi C = np.einsum('i,ijkl->ijkl', exparray, Y_mul_innerprod) # multiply the gradient of the gaussian with the other terms gradexp_mul_y_inner = np.einsum('in,ijkl->ijkln', gradexparray, Y_mul_innerprod) # add gradient of inner product and spherical harmonic terms gradHarmonics_mul_gaussian = np.einsum('ijkln,i->ijkln', dY_mul_innerprod+Y_mul_dinnerprod, exparray) dC = gradexp_mul_y_inner + gradHarmonics_mul_gaussian dC *= cutoff_array[:, np.newaxis, np.newaxis, np.newaxis, np.newaxis] dC += np.einsum('in,ijkl->ijkln', grad_cutoff_array, C) C *= cutoff_array[:, np.newaxis, np.newaxis, np.newaxis] return C, dC
if __name__ == "__main__": from optparse import OptionParser from ase.io import read import time # ---------------------- Options ------------------------ parser = OptionParser() parser.add_option("-c", "--crystal", dest="structure", help="crystal from file, cif or poscar, REQUIRED", metavar="crystal") parser.add_option("-r", "--rcut", dest="rcut", default=3.0, type=float, help="cutoff for neighbor calcs, default: 3.0" ) parser.add_option("-l", "--lmax", dest="lmax", default=2, type=int, help="lmax, default: 2" ) parser.add_option("-n", "--nmax", dest="nmax", default=2, type=int, help="nmax, default: 2" ) parser.add_option("-a", "--alpha", dest="alpha", default=2.0, type=float, help="cutoff for neighbor calcs, default: 2.0" ) parser.add_option("-f", dest="der", default=True, action='store_false',help='derivative flag') (options, args) = parser.parse_args() if options.structure is None: from ase.build import bulk test = bulk('Si', 'diamond', a=5.459, cubic=True) else: test = read(options.structure, format='vasp') cell = test.get_cell() cell[0,1] += 0.5 test.set_cell(cell) lmax = options.lmax nmax = options.nmax rcut = options.rcut alpha = options.alpha der = options.der start1 = time.time() f = SO3(nmax=nmax, lmax=lmax, rcut=rcut, alpha=alpha) start2 = time.time() print(f) p = f.compute_p(test, atom_ids=[0]); print('from P', p) x = f.calculate(test, derivative=der) print('x', x['x']) dp, p = f.compute_dpdr(test); print('from dP', p) dp, p = f.compute_dpdr_5d(test); print('from dP5d', p) (x, spg, wps) = ([13.45493847, 0.98, 0.04, 0.86, 0.08, 0.94, 0.14, 0.22, 0.66, 0.22, 0.74, 0.28, 0.2, 0.18, 0.68], 226, ['192j', '192j', '192j', '96i', '192j']) from pyxtal import pyxtal xtal = pyxtal() xtal.from_spg_wps_rep(spg, wps, x, ['C']*len(wps)) atoms = xtal.to_ase() p1 = f.compute_p(atoms)[0]; print("from ase", p1) p2 = f.compute_p(atoms, atom_ids=[0])[0]; print("from self", p2) print(np.sum((p1-p2)**2)) assert(np.sum((p1-p2)**2)<1e-4)