#!/usr/bin/env python
# Copyright 2014-2020 The PySCF Developers. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# Author: Qiming Sun <osirpt.sun@gmail.com>
#
'''
CCSD analytical nuclear gradients
'''
import ctypes
import numpy
from functools import reduce
from pyscf import lib
from pyscf import gto
from pyscf.lib import logger
from pyscf.cc import ccsd
from pyscf.cc import _ccsd
from pyscf.cc import ccsd_rdm
from pyscf.ao2mo import _ao2mo
from pyscf.scf import cphf
from pyscf.grad import rhf as rhf_grad
from pyscf.grad.mp2 import _shell_prange, _index_frozen_active
#
# Note: only works with canonical orbitals
# Non-canonical formula refers to JCP 95, 2639 (1991); DOI:10.1063/1.460916
#
[docs]
def grad_elec(cc_grad, t1=None, t2=None, l1=None, l2=None, eris=None, atmlst=None,
d1=None, d2=None, verbose=logger.INFO):
mycc = cc_grad.base
if eris is not None:
if abs(eris.fock - numpy.diag(eris.fock.diagonal())).max() > 1e-3:
raise RuntimeError('CCSD gradients does not support NHF (non-canonical HF)')
if t1 is None: t1 = mycc.t1
if t2 is None: t2 = mycc.t2
if l1 is None: l1 = mycc.l1
if l2 is None: l2 = mycc.l2
log = logger.new_logger(mycc, verbose)
time0 = logger.process_clock(), logger.perf_counter()
log.debug('Build ccsd rdm1 intermediates')
if d1 is None:
d1 = ccsd_rdm._gamma1_intermediates(mycc, t1, t2, l1, l2)
doo, dov, dvo, dvv = d1
time1 = log.timer_debug1('rdm1 intermediates', *time0)
log.debug('Build ccsd rdm2 intermediates')
fdm2 = lib.H5TmpFile()
if d2 is None:
d2 = ccsd_rdm._gamma2_outcore(mycc, t1, t2, l1, l2, fdm2, True)
time1 = log.timer_debug1('rdm2 intermediates', *time1)
mol = cc_grad.mol
mo_coeff = mycc.mo_coeff
mo_energy = mycc._scf.mo_energy
nao, nmo = mo_coeff.shape
nocc = numpy.count_nonzero(mycc.mo_occ > 0)
with_frozen = not ((mycc.frozen is None)
or (isinstance(mycc.frozen, (int, numpy.integer)) and mycc.frozen == 0)
or (len(mycc.frozen) == 0))
OA, VA, OF, VF = _index_frozen_active(mycc.get_frozen_mask(), mycc.mo_occ)
log.debug('symmetrized rdm2 and MO->AO transformation')
# Roughly, dm2*2 is computed in _rdm2_mo2ao
mo_active = mo_coeff[:,numpy.hstack((OA,VA))]
_rdm2_mo2ao(mycc, d2, mo_active, fdm2) # transform the active orbitals
time1 = log.timer_debug1('MO->AO transformation', *time1)
hf_dm1 = mycc._scf.make_rdm1(mycc.mo_coeff, mycc.mo_occ)
if atmlst is None:
atmlst = range(mol.natm)
offsetdic = mol.offset_nr_by_atom()
diagidx = numpy.arange(nao)
diagidx = diagidx*(diagidx+1)//2 + diagidx
de = numpy.zeros((len(atmlst),3))
Imat = numpy.zeros((nao,nao))
vhf1 = fdm2.create_dataset('vhf1', (len(atmlst),3,nao,nao), 'f8')
# 2e AO integrals dot 2pdm
max_memory = max(0, mycc.max_memory - lib.current_memory()[0])
blksize = max(1, int(max_memory*.9e6/8/(nao**3*2.5)))
for k, ia in enumerate(atmlst):
shl0, shl1, p0, p1 = offsetdic[ia]
ip1 = p0
vhf = numpy.zeros((3,nao,nao))
for b0, b1, nf in _shell_prange(mol, shl0, shl1, blksize):
ip0, ip1 = ip1, ip1 + nf
dm2buf = _load_block_tril(fdm2['dm2'], ip0, ip1, nao)
dm2buf[:,:,diagidx] *= .5
shls_slice = (b0,b1,0,mol.nbas,0,mol.nbas,0,mol.nbas)
eri0 = mol.intor('int2e', aosym='s2kl', shls_slice=shls_slice)
Imat += lib.einsum('ipx,iqx->pq', eri0.reshape(nf,nao,-1), dm2buf)
eri0 = None
eri1 = mol.intor('int2e_ip1', comp=3, aosym='s2kl',
shls_slice=shls_slice).reshape(3,nf,nao,-1)
de[k] -= numpy.einsum('xijk,ijk->x', eri1, dm2buf) * 2
dm2buf = None
# HF part
for i in range(3):
eri1tmp = lib.unpack_tril(eri1[i].reshape(nf*nao,-1))
eri1tmp = eri1tmp.reshape(nf,nao,nao,nao)
vhf[i] += numpy.einsum('ijkl,ij->kl', eri1tmp, hf_dm1[ip0:ip1])
vhf[i] -= numpy.einsum('ijkl,il->kj', eri1tmp, hf_dm1[ip0:ip1]) * .5
vhf[i,ip0:ip1] += numpy.einsum('ijkl,kl->ij', eri1tmp, hf_dm1)
vhf[i,ip0:ip1] -= numpy.einsum('ijkl,jk->il', eri1tmp, hf_dm1) * .5
eri1 = eri1tmp = None
vhf1[k] = vhf
log.debug('2e-part grad of atom %d %s = %s', ia, mol.atom_symbol(ia), de[k])
time1 = log.timer_debug1('2e-part grad of atom %d'%ia, *time1)
Imat = reduce(numpy.dot, (mo_coeff.T, Imat, mycc._scf.get_ovlp(), mo_coeff)) * -1
dm1mo = numpy.zeros((nmo,nmo))
if with_frozen:
dco = Imat[OF[:,None],OA] / (mo_energy[OF,None] - mo_energy[OA])
dfv = Imat[VF[:,None],VA] / (mo_energy[VF,None] - mo_energy[VA])
dm1mo[OA[:,None],OA] = doo + doo.T
dm1mo[OF[:,None],OA] = dco
dm1mo[OA[:,None],OF] = dco.T
dm1mo[VA[:,None],VA] = dvv + dvv.T
dm1mo[VF[:,None],VA] = dfv
dm1mo[VA[:,None],VF] = dfv.T
else:
dm1mo[:nocc,:nocc] = doo + doo.T
dm1mo[nocc:,nocc:] = dvv + dvv.T
dm1 = reduce(numpy.dot, (mo_coeff, dm1mo, mo_coeff.T))
vhf = mycc._scf.get_veff(mycc.mol, dm1) * 2
Xvo = reduce(numpy.dot, (mo_coeff[:,nocc:].T, vhf, mo_coeff[:,:nocc]))
Xvo+= Imat[:nocc,nocc:].T - Imat[nocc:,:nocc]
dm1mo += _response_dm1(mycc, Xvo, eris)
time1 = log.timer_debug1('response_rdm1 intermediates', *time1)
Imat[nocc:,:nocc] = Imat[:nocc,nocc:].T
im1 = reduce(numpy.dot, (mo_coeff, Imat, mo_coeff.T))
time1 = log.timer_debug1('response_rdm1', *time1)
log.debug('h1 and JK1')
# Initialize hcore_deriv with the underlying SCF object because some
# extensions (e.g. QM/MM, solvent) modifies the SCF object only.
mf_grad = cc_grad.base._scf.nuc_grad_method()
hcore_deriv = mf_grad.hcore_generator(mol)
s1 = mf_grad.get_ovlp(mol)
zeta = lib.direct_sum('i+j->ij', mo_energy, mo_energy) * .5
zeta[nocc:,:nocc] = mo_energy[:nocc]
zeta[:nocc,nocc:] = mo_energy[:nocc].reshape(-1,1)
zeta = reduce(numpy.dot, (mo_coeff, zeta*dm1mo, mo_coeff.T))
dm1 = reduce(numpy.dot, (mo_coeff, dm1mo, mo_coeff.T))
p1 = numpy.dot(mo_coeff[:,:nocc], mo_coeff[:,:nocc].T)
vhf_s1occ = reduce(numpy.dot, (p1, mycc._scf.get_veff(mol, dm1+dm1.T), p1))
time1 = log.timer_debug1('h1 and JK1', *time1)
# Hartree-Fock part contribution
dm1p = hf_dm1 + dm1*2
dm1 += hf_dm1
zeta += rhf_grad.make_rdm1e(mo_energy, mo_coeff, mycc.mo_occ)
for k, ia in enumerate(atmlst):
shl0, shl1, p0, p1 = offsetdic[ia]
# s[1] dot I, note matrix im1 is not hermitian
de[k] += numpy.einsum('xij,ij->x', s1[:,p0:p1], im1[p0:p1])
de[k] += numpy.einsum('xji,ij->x', s1[:,p0:p1], im1[:,p0:p1])
# h[1] \dot DM, contribute to f1
h1ao = hcore_deriv(ia)
de[k] += numpy.einsum('xij,ji->x', h1ao, dm1)
# -s[1]*e \dot DM, contribute to f1
de[k] -= numpy.einsum('xij,ij->x', s1[:,p0:p1], zeta[p0:p1] )
de[k] -= numpy.einsum('xji,ij->x', s1[:,p0:p1], zeta[:,p0:p1])
# -vhf[s_ij[1]], contribute to f1, *2 for s1+s1.T
de[k] -= numpy.einsum('xij,ij->x', s1[:,p0:p1], vhf_s1occ[p0:p1]) * 2
de[k] -= numpy.einsum('xij,ij->x', vhf1[k], dm1p)
log.timer('%s gradients' % mycc.__class__.__name__, *time0)
return de
[docs]
def as_scanner(grad_cc):
'''Generating a nuclear gradients scanner/solver (for geometry optimizer).
The returned solver is a function. This function requires one argument
"mol" as input and returns total CCSD energy.
The solver will automatically use the results of last calculation as the
initial guess of the new calculation. All parameters assigned in the
CCSD and the underlying SCF objects (conv_tol, max_memory etc) are
automatically applied in the solver.
Note scanner has side effects. It may change many underlying objects
(_scf, with_df, with_x2c, ...) during calculation.
Examples::
>>> from pyscf import gto, scf, cc
>>> mol = gto.M(atom='H 0 0 0; F 0 0 1')
>>> cc_scanner = cc.CCSD(scf.RHF(mol)).nuc_grad_method().as_scanner()
>>> e_tot, grad = cc_scanner(gto.M(atom='H 0 0 0; F 0 0 1.1'))
>>> e_tot, grad = cc_scanner(gto.M(atom='H 0 0 0; F 0 0 1.5'))
'''
from pyscf import gto
if isinstance(grad_cc, lib.GradScanner):
return grad_cc
logger.info(grad_cc, 'Create scanner for %s', grad_cc.__class__)
name = grad_cc.__class__.__name__ + CCSD_GradScanner.__name_mixin__
return lib.set_class(CCSD_GradScanner(grad_cc),
(CCSD_GradScanner, grad_cc.__class__), name)
[docs]
class CCSD_GradScanner(lib.GradScanner):
def __init__(self, g):
lib.GradScanner.__init__(self, g)
def __call__(self, mol_or_geom, **kwargs):
if isinstance(mol_or_geom, gto.MoleBase):
assert mol_or_geom.__class__ == gto.Mole
mol = mol_or_geom
else:
mol = self.mol.set_geom_(mol_or_geom, inplace=False)
cc = self.base
if cc.t2 is not None:
last_size = cc.vector_size()
else:
last_size = 0
self.reset(mol)
mf_scanner = cc._scf
mf_scanner(mol)
cc.mo_coeff = mf_scanner.mo_coeff
cc.mo_occ = mf_scanner.mo_occ
if last_size != cc.vector_size():
cc.t1 = cc.t2 = cc.l1 = cc.l2 = None
eris = cc.ao2mo(cc.mo_coeff)
# Update cc.t1 and cc.t2
cc.kernel(t1=cc.t1, t2=cc.t2, eris=eris)
# Update cc.l1 and cc.l2
cc.solve_lambda(l1=cc.l1, l2=cc.l2, eris=eris)
de = self.kernel(cc.t1, cc.t2, cc.l1, cc.l2, eris=eris, **kwargs)
return cc.e_tot, de
@property
def converged(self):
cc = self.base
return all((cc._scf.converged, cc.converged, cc.converged_lambda))
def _response_dm1(mycc, Xvo, eris=None):
nvir, nocc = Xvo.shape
nmo = nocc + nvir
with_frozen = not ((mycc.frozen is None)
or (isinstance(mycc.frozen, (int, numpy.integer)) and mycc.frozen == 0)
or (len(mycc.frozen) == 0))
if eris is None or with_frozen:
mo_energy = mycc._scf.mo_energy
mo_occ = mycc.mo_occ
mo_coeff = mycc.mo_coeff
def fvind(x):
x = x.reshape(Xvo.shape)
dm = reduce(numpy.dot, (mo_coeff[:,nocc:], x, mo_coeff[:,:nocc].T))
v = mycc._scf.get_veff(mycc.mol, dm + dm.T)
v = reduce(numpy.dot, (mo_coeff[:,nocc:].T, v, mo_coeff[:,:nocc]))
return v * 2
else:
mo_energy = eris.mo_energy
mo_occ = numpy.zeros_like(mo_energy)
mo_occ[:nocc] = 2
ovvo = numpy.empty((nocc,nvir,nvir,nocc))
for i in range(nocc):
ovvo[i] = eris.ovvo[i]
ovvo[i] = ovvo[i] * 4 - ovvo[i].transpose(1,0,2)
ovvo[i]-= eris.oovv[i].transpose(2,1,0)
def fvind(x):
return numpy.einsum('iabj,bj->ai', ovvo, x.reshape(Xvo.shape))
dvo = cphf.solve(fvind, mo_energy, mo_occ, Xvo, max_cycle=30)[0]
dm1 = numpy.zeros((nmo,nmo))
dm1[nocc:,:nocc] = dvo
dm1[:nocc,nocc:] = dvo.T
return dm1
def _rdm2_mo2ao(mycc, d2, mo_coeff, fsave=None):
# dm2 = ccsd_rdm._make_rdm2(mycc, None, d2, with_dm1=False)
# dm2 = numpy.einsum('pi,ijkl->pjkl', mo_coeff, dm2)
# dm2 = numpy.einsum('pj,ijkl->ipkl', mo_coeff, dm2)
# dm2 = numpy.einsum('pk,ijkl->ijpl', mo_coeff, dm2)
# dm2 = numpy.einsum('pl,ijkl->ijkp', mo_coeff, dm2)
# dm2 = dm2 + dm2.transpose(1,0,2,3)
# dm2 = dm2 + dm2.transpose(0,1,3,2)
# return ao2mo.restore(4, dm2*.5, nmo)
log = logger.Logger(mycc.stdout, mycc.verbose)
time1 = logger.process_clock(), logger.perf_counter()
if fsave is None:
incore = True
fsave = lib.H5TmpFile()
else:
incore = False
dovov, dvvvv, doooo, doovv, dovvo, dvvov, dovvv, dooov = d2
nocc, nvir = dovov.shape[:2]
mo_coeff = numpy.asarray(mo_coeff, order='F')
nao, nmo = mo_coeff.shape
nao_pair = nao * (nao+1) // 2
nvir_pair = nvir * (nvir+1) //2
fdrv = _ao2mo.libao2mo.AO2MOnr_e2_drv
ftrans = _ao2mo.libao2mo.AO2MOtranse2_nr_s1
fmm = _ccsd.libcc.CCmmm_transpose_sum
pao_loc = ctypes.POINTER(ctypes.c_void_p)()
def _trans(vin, orbs_slice, out=None):
nrow = vin.shape[0]
if out is None:
out = numpy.empty((nrow,nao_pair))
fdrv(ftrans, fmm,
out.ctypes.data_as(ctypes.c_void_p),
vin.ctypes.data_as(ctypes.c_void_p),
mo_coeff.ctypes.data_as(ctypes.c_void_p),
ctypes.c_int(nrow), ctypes.c_int(nao),
(ctypes.c_int*4)(*orbs_slice), pao_loc, ctypes.c_int(0))
return out
fswap = lib.H5TmpFile()
max_memory = mycc.max_memory - lib.current_memory()[0]
blksize = int(max_memory*1e6/8/(nao_pair+nmo**2))
blksize = min(nvir_pair, max(ccsd.BLKMIN, blksize))
chunks_vv = (int(min(blksize,4e8/blksize)), blksize)
fswap.create_dataset('v', (nao_pair,nvir_pair), 'f8', chunks=chunks_vv)
for p0, p1 in lib.prange(0, nvir_pair, blksize):
fswap['v'][:,p0:p1] = _trans(lib.unpack_tril(_cp(dvvvv[p0:p1])),
(nocc,nmo,nocc,nmo)).T
time1 = log.timer_debug1('_rdm2_mo2ao pass 1', *time1)
# transform dm2_ij to get lower triangular (dm2+dm2.transpose(0,1,3,2))
blksize = int(max_memory*1e6/8/(nao_pair+nmo**2))
blksize = min(nao_pair, max(ccsd.BLKMIN, blksize))
fswap.create_dataset('o', (nmo,nocc,nao_pair), 'f8', chunks=(nocc,nocc,blksize))
buf1 = numpy.zeros((nocc,nocc,nmo,nmo))
buf1[:,:,:nocc,:nocc] = doooo
buf1[:,:,nocc:,nocc:] = _cp(doovv)
buf1 = _trans(buf1.reshape(nocc**2,-1), (0,nmo,0,nmo))
fswap['o'][:nocc] = buf1.reshape(nocc,nocc,nao_pair)
dovoo = numpy.asarray(dooov).transpose(2,3,0,1)
for p0, p1 in lib.prange(nocc, nmo, nocc):
buf1 = numpy.zeros((nocc,p1-p0,nmo,nmo))
buf1[:,:,:nocc,:nocc] = dovoo[:,p0-nocc:p1-nocc]
buf1[:,:,nocc:,:nocc] = dovvo[:,p0-nocc:p1-nocc]
buf1[:,:,:nocc,nocc:] = dovov[:,p0-nocc:p1-nocc]
buf1[:,:,nocc:,nocc:] = dovvv[:,p0-nocc:p1-nocc]
buf1 = buf1.transpose(1,0,3,2).reshape((p1-p0)*nocc,-1)
buf1 = _trans(buf1, (0,nmo,0,nmo))
fswap['o'][p0:p1] = buf1.reshape(p1-p0,nocc,nao_pair)
time1 = log.timer_debug1('_rdm2_mo2ao pass 2', *time1)
dovoo = buf1 = None
# transform dm2_kl then dm2 + dm2.transpose(2,3,0,1)
gsave = fsave.create_dataset('dm2', (nao_pair,nao_pair), 'f8', chunks=chunks_vv)
for p0, p1 in lib.prange(0, nao_pair, blksize):
buf1 = numpy.zeros((p1-p0,nmo,nmo))
buf1[:,nocc:,nocc:] = lib.unpack_tril(_cp(fswap['v'][p0:p1]))
buf1[:,:,:nocc] = fswap['o'][:,:,p0:p1].transpose(2,0,1)
buf2 = _trans(buf1, (0,nmo,0,nmo))
if p0 > 0:
buf1 = _cp(gsave[:p0,p0:p1])
buf1[:p0,:p1-p0] += buf2[:p1-p0,:p0].T
buf2[:p1-p0,:p0] = buf1[:p0,:p1-p0].T
gsave[:p0,p0:p1] = buf1
lib.transpose_sum(buf2[:,p0:p1], inplace=True)
gsave[p0:p1] = buf2
time1 = log.timer_debug1('_rdm2_mo2ao pass 3', *time1)
if incore:
return fsave['dm2'][:]
else:
return fsave
#
# .
# . .
# ----+ -----------
# ----|-+ => -----------
# . . | | .
# . . | | . .
#
def _load_block_tril(h5dat, row0, row1, nao, out=None):
nao_pair = nao * (nao+1) // 2
if out is None:
out = numpy.ndarray((row1-row0,nao,nao_pair))
dat = h5dat[row0*(row0+1)//2:row1*(row1+1)//2]
p1 = 0
for i in range(row0, row1):
p0, p1 = p1, p1 + i+1
out[i-row0,:i+1] = dat[p0:p1]
for j in range(row0, i):
out[j-row0,i] = out[i-row0,j]
for i in range(row1, nao):
i2 = i*(i+1)//2
out[:,i] = h5dat[i2+row0:i2+row1]
return out
def _cp(a):
return numpy.asarray(a, order='C')
[docs]
class Gradients(rhf_grad.GradientsBase):
grad_elec = grad_elec
[docs]
def kernel(self, t1=None, t2=None, l1=None, l2=None, eris=None,
atmlst=None, verbose=None):
log = logger.new_logger(self, verbose)
mycc = self.base
if t1 is None: t1 = mycc.t1
if t2 is None: t2 = mycc.t2
if l1 is None: l1 = mycc.l1
if l2 is None: l2 = mycc.l2
if eris is None:
eris = mycc.ao2mo()
if t1 is None or t2 is None:
t1, t2 = mycc.kernel(eris=eris)[1:]
if l1 is None or l2 is None:
l1, l2 = mycc.solve_lambda(eris=eris)
if atmlst is None:
atmlst = self.atmlst
else:
self.atmlst = atmlst
de = self.grad_elec(t1, t2, l1, l2, eris, atmlst, verbose=log)
self.de = de + self.grad_nuc(atmlst=atmlst)
if self.mol.symmetry:
self.de = self.symmetrize(self.de, atmlst)
self._finalize()
return self.de
# Calling the underlying SCF nuclear gradients because it may be modified
# by external modules (e.g. QM/MM, solvent)
[docs]
def grad_nuc(self, mol=None, atmlst=None):
mf_grad = self.base._scf.nuc_grad_method()
return mf_grad.grad_nuc(mol, atmlst)
as_scanner = as_scanner
to_gpu = lib.to_gpu
Grad = Gradients
ccsd.CCSD.Gradients = lib.class_as_method(Gradients)
