def kernel(mycc,
t1=None,
t2=None,
l1=None,
l2=None,
eris=None,
atmlst=None,
mf_grad=None,
d1=None,
d2=None,
verbose=logger.INFO):
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
if mf_grad is None: mf_grad = mycc._scf.nuc_grad_method()
log = logger.new_logger(mycc, verbose)
time0 = time.clock(), time.time()
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 = mycc.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 mycc.frozen is 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')
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 += mf_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)
de += mf_grad.grad_nuc(mol, atmlst)
log.timer('%s gradients' % mycc.__class__.__name__, *time0)
return de
def kernel(mc,
mo_coeff=None,
ci=None,
atmlst=None,
mf_grad=None,
verbose=None):
if mo_coeff is None: mo_coeff = mc.mo_coeff
if ci is None: ci = mc.ci
if mf_grad is None: mf_grad = mc._scf.nuc_grad_method()
if mc.frozen is not None:
raise NotImplementedError
mol = mc.mol
ncore = mc.ncore
ncas = mc.ncas
nocc = ncore + ncas
nelecas = mc.nelecas
nao, nmo = mo_coeff.shape
nao_pair = nao * (nao + 1) // 2
mo_occ = mo_coeff[:, :nocc]
mo_core = mo_coeff[:, :ncore]
mo_cas = mo_coeff[:, ncore:nocc]
casdm1, casdm2 = mc.fcisolver.make_rdm12(ci, ncas, nelecas)
# gfock = Generalized Fock, Adv. Chem. Phys., 69, 63
dm_core = numpy.dot(mo_core, mo_core.T) * 2
dm_cas = reduce(numpy.dot, (mo_cas, casdm1, mo_cas.T))
aapa = ao2mo.kernel(mol, (mo_cas, mo_cas, mo_occ, mo_cas), compact=False)
aapa = aapa.reshape(ncas, ncas, nocc, ncas)
vj, vk = mc._scf.get_jk(mol, (dm_core, dm_cas))
h1 = mc.get_hcore()
vhf_c = vj[0] - vk[0] * .5
vhf_a = vj[1] - vk[1] * .5
gfock = reduce(numpy.dot, (mo_occ.T, h1 + vhf_c + vhf_a, mo_occ)) * 2
gfock[:, ncore:nocc] = reduce(numpy.dot,
(mo_occ.T, h1 + vhf_c, mo_cas, casdm1))
gfock[:, ncore:nocc] += numpy.einsum('uviw,vuwt->it', aapa, casdm2)
dme0 = reduce(numpy.dot, (mo_occ, (gfock + gfock.T) * .5, mo_occ.T))
aapa = vj = vk = vhf_c = vhf_a = h1 = gfock = None
dm1 = dm_core + dm_cas
vhf1c, vhf1a = mf_grad.get_veff(mol, (dm_core, dm_cas))
hcore_deriv = mf_grad.hcore_generator(mol)
s1 = mf_grad.get_ovlp(mol)
diag_idx = numpy.arange(nao)
diag_idx = diag_idx * (diag_idx + 1) // 2 + diag_idx
casdm2_cc = casdm2 + casdm2.transpose(0, 1, 3, 2)
dm2buf = ao2mo._ao2mo.nr_e2(casdm2_cc.reshape(ncas**2, ncas**2), mo_cas.T,
(0, nao, 0, nao)).reshape(ncas**2, nao, nao)
dm2buf = lib.pack_tril(dm2buf)
dm2buf[:, diag_idx] *= .5
dm2buf = dm2buf.reshape(ncas, ncas, nao_pair)
casdm2 = casdm2_cc = None
if atmlst is None:
atmlst = range(mol.natm)
aoslices = mol.aoslice_by_atom()
de = numpy.zeros((len(atmlst), 3))
max_memory = mc.max_memory - lib.current_memory()[0]
blksize = int(max_memory * .9e6 / 8 /
((aoslices[:, 3] - aoslices[:, 2]).max() * nao_pair))
blksize = min(nao, max(2, blksize))
for k, ia in enumerate(atmlst):
shl0, shl1, p0, p1 = aoslices[ia]
h1ao = hcore_deriv(ia)
de[k] += numpy.einsum('xij,ij->x', h1ao, dm1)
de[k] -= numpy.einsum('xij,ij->x', s1[:, p0:p1], dme0[p0:p1]) * 2
q1 = 0
for b0, b1, nf in _shell_prange(mol, 0, mol.nbas, blksize):
q0, q1 = q1, q1 + nf
dm2_ao = lib.einsum('ijw,pi,qj->pqw', dm2buf, mo_cas[p0:p1],
mo_cas[q0:q1])
shls_slice = (shl0, shl1, b0, b1, 0, mol.nbas, 0, mol.nbas)
eri1 = mol.intor('int2e_ip1',
comp=3,
aosym='s2kl',
shls_slice=shls_slice).reshape(
3, p1 - p0, nf, nao_pair)
de[k] -= numpy.einsum('xijw,ijw->x', eri1, dm2_ao) * 2
eri1 = None
de[k] += numpy.einsum('xij,ij->x', vhf1c[:, p0:p1], dm1[p0:p1]) * 2
de[k] += numpy.einsum('xij,ij->x', vhf1a[:, p0:p1], dm_core[p0:p1]) * 2
de += rhf_grad.grad_nuc(mol, atmlst)
return de
def kernel(mc,
mo_coeff=None,
ci=None,
atmlst=None,
mf_grad=None,
verbose=None):
if mo_coeff is None: mo_coeff = mc._scf.mo_coeff
if ci is None: ci = mc.ci
if mf_grad is None: mf_grad = mc._scf.nuc_grad_method()
assert (isinstance(ci, numpy.ndarray))
mol = mc.mol
ncore = mc.ncore
ncas = mc.ncas
nocc = ncore + ncas
nelecas = mc.nelecas
nao, nmo = mo_coeff.shape
nao_pair = nao * (nao + 1) // 2
mo_energy = mc._scf.mo_energy
mo_occ = mo_coeff[:, :nocc]
mo_core = mo_coeff[:, :ncore]
mo_cas = mo_coeff[:, ncore:nocc]
neleca, nelecb = mol.nelec
assert (neleca == nelecb)
orbo = mo_coeff[:, :neleca]
orbv = mo_coeff[:, neleca:]
casdm1, casdm2 = mc.fcisolver.make_rdm12(ci, ncas, nelecas)
dm_core = numpy.dot(mo_core, mo_core.T) * 2
dm_cas = reduce(numpy.dot, (mo_cas, casdm1, mo_cas.T))
aapa = ao2mo.kernel(mol, (mo_cas, mo_cas, mo_coeff, mo_cas), compact=False)
aapa = aapa.reshape(ncas, ncas, nmo, ncas)
vj, vk = mc._scf.get_jk(mol, (dm_core, dm_cas))
h1 = mc.get_hcore()
vhf_c = vj[0] - vk[0] * .5
vhf_a = vj[1] - vk[1] * .5
# Imat = h1_{pi} gamma1_{iq} + h2_{pijk} gamma_{iqkj}
Imat = numpy.zeros((nmo, nmo))
Imat[:, :nocc] = reduce(numpy.dot,
(mo_coeff.T, h1 + vhf_c + vhf_a, mo_occ)) * 2
Imat[:, ncore:nocc] = reduce(numpy.dot,
(mo_coeff.T, h1 + vhf_c, mo_cas, casdm1))
Imat[:, ncore:nocc] += lib.einsum('uviw,vuwt->it', aapa, casdm2)
aapa = vj = vk = vhf_c = vhf_a = h1 = None
ee = mo_energy[:, None] - mo_energy
zvec = numpy.zeros_like(Imat)
zvec[:ncore,
ncore:neleca] = Imat[:ncore, ncore:neleca] / -ee[:ncore, ncore:neleca]
zvec[ncore:neleca, :ncore] = Imat[
ncore:neleca, :ncore] / -ee[ncore:neleca, :ncore]
zvec[nocc:,
neleca:nocc] = Imat[nocc:, neleca:nocc] / -ee[nocc:, neleca:nocc]
zvec[neleca:nocc,
nocc:] = Imat[neleca:nocc, nocc:] / -ee[neleca:nocc, nocc:]
zvec_ao = reduce(numpy.dot, (mo_coeff, zvec + zvec.T, mo_coeff.T))
vhf = mc._scf.get_veff(mol, zvec_ao) * 2
xvo = reduce(numpy.dot, (orbv.T, vhf, orbo))
xvo += Imat[neleca:, :neleca] - Imat[:neleca, neleca:].T
def fvind(x):
x = x.reshape(xvo.shape)
dm = reduce(numpy.dot, (orbv, x, orbo.T))
v = mc._scf.get_veff(mol, dm + dm.T)
v = reduce(numpy.dot, (orbv.T, v, orbo))
return v * 2
dm1resp = cphf.solve(fvind, mo_energy, mc._scf.mo_occ, xvo,
max_cycle=30)[0]
zvec[neleca:, :neleca] = dm1resp
zeta = numpy.einsum('ij,j->ij', zvec, mo_energy)
zeta = reduce(numpy.dot, (mo_coeff, zeta, mo_coeff.T))
zvec_ao = reduce(numpy.dot, (mo_coeff, zvec + zvec.T, mo_coeff.T))
p1 = numpy.dot(mo_coeff[:, :neleca], mo_coeff[:, :neleca].T)
vhf_s1occ = reduce(numpy.dot, (p1, mc._scf.get_veff(mol, zvec_ao), p1))
Imat[:ncore, ncore:neleca] = 0
Imat[ncore:neleca, :ncore] = 0
Imat[nocc:, neleca:nocc] = 0
Imat[neleca:nocc, nocc:] = 0
Imat[neleca:, :neleca] = Imat[:neleca, neleca:].T
im1 = reduce(numpy.dot, (mo_coeff, Imat, mo_coeff.T))
casci_dm1 = dm_core + dm_cas
hf_dm1 = mc._scf.make_rdm1(mo_coeff, mc._scf.mo_occ)
hcore_deriv = mf_grad.hcore_generator(mol)
s1 = mf_grad.get_ovlp(mol)
diag_idx = numpy.arange(nao)
diag_idx = diag_idx * (diag_idx + 1) // 2 + diag_idx
casdm2_cc = casdm2 + casdm2.transpose(0, 1, 3, 2)
dm2buf = ao2mo._ao2mo.nr_e2(casdm2_cc.reshape(ncas**2, ncas**2), mo_cas.T,
(0, nao, 0, nao)).reshape(ncas**2, nao, nao)
dm2buf = lib.pack_tril(dm2buf)
dm2buf[:, diag_idx] *= .5
dm2buf = dm2buf.reshape(ncas, ncas, nao_pair)
casdm2 = casdm2_cc = None
if atmlst is None:
atmlst = range(mol.natm)
aoslices = mol.aoslice_by_atom()
de = numpy.zeros((len(atmlst), 3))
max_memory = mc.max_memory - lib.current_memory()[0]
blksize = int(max_memory * .9e6 / 8 /
((aoslices[:, 3] - aoslices[:, 2]).max() * nao_pair))
blksize = min(nao, max(2, blksize))
for k, ia in enumerate(atmlst):
shl0, shl1, p0, p1 = aoslices[ia]
h1ao = hcore_deriv(ia)
de[k] += numpy.einsum('xij,ij->x', h1ao, casci_dm1)
de[k] += numpy.einsum('xij,ij->x', h1ao, zvec_ao)
vhf1 = numpy.zeros((3, nao, nao))
q1 = 0
for b0, b1, nf in _shell_prange(mol, 0, mol.nbas, blksize):
q0, q1 = q1, q1 + nf
dm2_ao = lib.einsum('ijw,pi,qj->pqw', dm2buf, mo_cas[p0:p1],
mo_cas[q0:q1])
shls_slice = (shl0, shl1, b0, b1, 0, mol.nbas, 0, mol.nbas)
eri1 = mol.intor('int2e_ip1',
comp=3,
aosym='s2kl',
shls_slice=shls_slice).reshape(
3, p1 - p0, nf, nao_pair)
de[k] -= numpy.einsum('xijw,ijw->x', eri1, dm2_ao) * 2
for i in range(3):
eri1tmp = lib.unpack_tril(eri1[i].reshape((p1 - p0) * nf, -1))
eri1tmp = eri1tmp.reshape(p1 - p0, nf, nao, nao)
de[k, i] -= numpy.einsum('ijkl,ij,kl', eri1tmp,
hf_dm1[p0:p1, q0:q1], zvec_ao) * 2
de[k, i] -= numpy.einsum('ijkl,kl,ij', eri1tmp, hf_dm1,
zvec_ao[p0:p1, q0:q1]) * 2
de[k, i] += numpy.einsum('ijkl,il,kj', eri1tmp, hf_dm1[p0:p1],
zvec_ao[q0:q1])
de[k, i] += numpy.einsum('ijkl,jk,il', eri1tmp, hf_dm1[q0:q1],
zvec_ao[p0:p1])
#:vhf1c, vhf1a = mf_grad.get_veff(mol, (dm_core, dm_cas))
#:de[k] += numpy.einsum('xij,ij->x', vhf1c[:,p0:p1], casci_dm1[p0:p1]) * 2
#:de[k] += numpy.einsum('xij,ij->x', vhf1a[:,p0:p1], dm_core[p0:p1]) * 2
de[k, i] -= numpy.einsum('ijkl,lk,ij', eri1tmp, dm_core[q0:q1],
casci_dm1[p0:p1]) * 2
de[k, i] += numpy.einsum('ijkl,jk,il', eri1tmp, dm_core[q0:q1],
casci_dm1[p0:p1])
de[k, i] -= numpy.einsum('ijkl,lk,ij', eri1tmp, dm_cas[q0:q1],
dm_core[p0:p1]) * 2
de[k, i] += numpy.einsum('ijkl,jk,il', eri1tmp, dm_cas[q0:q1],
dm_core[p0:p1])
eri1 = eri1tmp = None
de[k] -= numpy.einsum('xij,ij->x', s1[:, p0:p1], im1[p0:p1])
de[k] -= numpy.einsum('xij,ji->x', s1[:, p0:p1], im1[:, p0:p1])
de[k] -= numpy.einsum('xij,ij->x', s1[:, p0:p1], zeta[p0:p1]) * 2
de[k] -= numpy.einsum('xij,ji->x', s1[:, p0:p1], zeta[:, p0:p1]) * 2
de[k] -= numpy.einsum('xij,ij->x', s1[:, p0:p1], vhf_s1occ[p0:p1]) * 2
de[k] -= numpy.einsum('xij,ji->x', s1[:, p0:p1], vhf_s1occ[:,
p0:p1]) * 2
de += rhf_grad.grad_nuc(mol, atmlst)
return de
def grad_elec(mc_grad, mo_coeff=None, ci=None, atmlst=None, verbose=None):
mc = mc_grad.base
if mo_coeff is None: mo_coeff = mc.mo_coeff
if ci is None: ci = mc.ci
if mc.frozen is not None:
raise NotImplementedError
time0 = time.clock(), time.time()
log = logger.new_logger(mc_grad, verbose)
mol = mc_grad.mol
ncore = mc.ncore
ncas = mc.ncas
nocc = ncore + ncas
nelecas = mc.nelecas
nao, nmo = mo_coeff.shape
nao_pair = nao * (nao+1) // 2
# Necessary kludge because gfock isn't zero in occ-virt space in SA-CASSCf
# Among many other potential applications!
if hasattr (mc, '_tag_gfock_ov_nonzero'):
if mc._tag_gfock_ov_nonzero:
nocc = nmo
mo_occ = mo_coeff[:,:nocc]
mo_core = mo_coeff[:,:ncore]
mo_cas = mo_coeff[:,ncore:ncore+ncas]
casdm1, casdm2 = mc.fcisolver.make_rdm12(ci, ncas, nelecas)
# gfock = Generalized Fock, Adv. Chem. Phys., 69, 63
dm_core = numpy.dot(mo_core, mo_core.T) * 2
dm_cas = reduce(numpy.dot, (mo_cas, casdm1, mo_cas.T))
# MRH flag: this is one of my kludges
# It would be better to just pass the ERIS object used in orbital optimization
# But I am too lazy at the moment
aapa = ao2mo.kernel(mol, (mo_cas, mo_cas, mo_occ, mo_cas), compact=False)
aapa = aapa.reshape(ncas,ncas,nocc,ncas)
vj, vk = mc._scf.get_jk(mol, (dm_core, dm_cas))
h1 = mc.get_hcore()
vhf_c = vj[0] - vk[0] * .5
vhf_a = vj[1] - vk[1] * .5
gfock = numpy.zeros ((nocc, nocc))
gfock[:,:ncore] = reduce(numpy.dot, (mo_occ.T, h1 + vhf_c + vhf_a, mo_core)) * 2
gfock[:,ncore:ncore+ncas] = reduce(numpy.dot, (mo_occ.T, h1 + vhf_c, mo_cas, casdm1))
gfock[:,ncore:ncore+ncas] += numpy.einsum('uviw,vuwt->it', aapa, casdm2)
dme0 = reduce(numpy.dot, (mo_occ, (gfock+gfock.T)*.5, mo_occ.T))
aapa = vj = vk = vhf_c = vhf_a = h1 = gfock = None
dm1 = dm_core + dm_cas
vj, vk = mc_grad.get_jk(mol, (dm_core, dm_cas))
vhf1c, vhf1a = vj - vk * .5
hcore_deriv = mc_grad.hcore_generator(mol)
s1 = mc_grad.get_ovlp(mol)
diag_idx = numpy.arange(nao)
diag_idx = diag_idx * (diag_idx+1) // 2 + diag_idx
casdm2_cc = casdm2 + casdm2.transpose(0,1,3,2)
dm2buf = ao2mo._ao2mo.nr_e2(casdm2_cc.reshape(ncas**2,ncas**2), mo_cas.T,
(0, nao, 0, nao)).reshape(ncas**2,nao,nao)
dm2buf = lib.pack_tril(dm2buf)
dm2buf[:,diag_idx] *= .5
dm2buf = dm2buf.reshape(ncas,ncas,nao_pair)
casdm2 = casdm2_cc = None
if atmlst is None:
atmlst = range(mol.natm)
aoslices = mol.aoslice_by_atom()
de = numpy.zeros((len(atmlst),3))
max_memory = mc_grad.max_memory - lib.current_memory()[0]
# MRH: this originally implied that the memory footprint would be max(p1-p0) * max(q1-q0) * nao_pair
# In fact, that's the size of dm2_ao AND EACH COMPONENT of the differentiated eris
# So the actual memory footprint is 4 times that!
blksize = int(max_memory*.9e6/8 / (4*(aoslices[:,3]-aoslices[:,2]).max()*nao_pair))
blksize = min(nao, max(2, blksize))
for k, ia in enumerate(atmlst):
shl0, shl1, p0, p1 = aoslices[ia]
h1ao = hcore_deriv(ia)
de[k] += numpy.einsum('xij,ij->x', h1ao, dm1)
de[k] -= numpy.einsum('xij,ij->x', s1[:,p0:p1], dme0[p0:p1]) * 2
q1 = 0
for b0, b1, nf in _shell_prange(mol, 0, mol.nbas, blksize):
q0, q1 = q1, q1 + nf
dm2_ao = lib.einsum('ijw,pi,qj->pqw', dm2buf, mo_cas[p0:p1], mo_cas[q0:q1])
shls_slice = (shl0,shl1,b0,b1,0,mol.nbas,0,mol.nbas)
eri1 = mol.intor('int2e_ip1', comp=3, aosym='s2kl',
shls_slice=shls_slice).reshape(3,p1-p0,nf,nao_pair)
de[k] -= numpy.einsum('xijw,ijw->x', eri1, dm2_ao) * 2
eri1 = None
de[k] += numpy.einsum('xij,ij->x', vhf1c[:,p0:p1], dm1[p0:p1]) * 2
de[k] += numpy.einsum('xij,ij->x', vhf1a[:,p0:p1], dm_core[p0:p1]) * 2
log.timer('CASSCF nuclear gradients', *time0)
return de
def kernel(mc,
mo_coeff=None,
ci=None,
atmlst=None,
mf_grad=None,
verbose=None):
if mo_coeff is None: mo_coeff = mc._scf.mo_coeff
if ci is None: ci = mc.ci
if mf_grad is None: mf_grad = mc._scf.nuc_grad_method()
mol = mc.mol
ncore = mc.ncore
ncas = mc.ncas
nocc = ncore + ncas
nelecas = mc.nelecas
nao, nmo = mo_coeff.shape
nao_pair = nao * (nao + 1) // 2
mo_energy = mc._scf.mo_energy
hcore_deriv = mf_grad.hcore_generator(mol)
s1 = mf_grad.get_ovlp(mol)
mo_occ = mo_coeff[:, :nocc]
mo_core = mo_coeff[:, :ncore]
mo_cas = mo_coeff[:, ncore:nocc]
casdm1, casdm2 = mc.fcisolver.make_rdm12(mc.ci, ncas, nelecas)
# gfock = Generalized Fock, Adv. Chem. Phys., 69, 63
dm_core = numpy.dot(mo_core, mo_core.T) * 2
dm_cas = reduce(numpy.dot, (mo_cas, casdm1, mo_cas.T))
aapa = ao2mo.kernel(mol, (mo_cas, mo_cas, mo_occ, mo_cas), compact=False)
aapa = aapa.reshape(ncas, ncas, nocc, ncas)
vj, vk = mc._scf.get_jk(mol, (dm_core, dm_cas))
h1 = mc.get_hcore()
vhf_c = vj[0] - vk[0] * .5
vhf_a = vj[1] - vk[1] * .5
gfock = reduce(numpy.dot, (mo_occ.T, h1 + vhf_c + vhf_a, mo_occ)) * 2
gfock[:, ncore:nocc] = reduce(numpy.dot,
(mo_occ.T, h1 + vhf_c, mo_cas, casdm1))
gfock[:, ncore:nocc] += numpy.einsum('uviw,vuwt->it', aapa, casdm2)
dme0 = reduce(numpy.dot, (mo_occ, (gfock + gfock.T) * .5, mo_occ.T))
aapa = vj = vk = vhf_c = vhf_a = h1 = gfock = None
dm1 = dm_core + dm_cas
vhf1c, vhf1a = mf_grad.get_veff(mol, (dm_core, dm_cas))
diag_idx = numpy.arange(nao)
diag_idx = diag_idx * (diag_idx + 1) // 2 + diag_idx
casdm2_cc = casdm2 + casdm2.transpose(0, 1, 3, 2)
dm2buf = ao2mo._ao2mo.nr_e2(casdm2_cc.reshape(ncas**2, ncas**2), mo_cas.T,
(0, nao, 0, nao)).reshape(ncas**2, nao, nao)
dm2buf = lib.pack_tril(dm2buf)
dm2buf[:, diag_idx] *= .5
dm2buf = dm2buf.reshape(ncas, ncas, nao_pair)
#casdm2 = casdm2_cc = None
atmlst = range(mol.natm)
aoslices = mol.aoslice_by_atom()
de = numpy.zeros((len(atmlst), 3))
max_memory = mc.max_memory - lib.current_memory()[0]
blksize = int(max_memory * .9e6 / 8 /
((aoslices[:, 3] - aoslices[:, 2]).max() * nao_pair))
blksize = min(nao, max(2, blksize))
for k, ia in enumerate(atmlst):
shl0, shl1, p0, p1 = aoslices[ia]
h1ao = hcore_deriv(ia)
de[k] += numpy.einsum('xij,ij->x', h1ao, dm1)
#de[k] -= numpy.einsum('xij,ij->x', s1[:,p0:p1], dme0[p0:p1]) * 2
q1 = 0
for b0, b1, nf in _shell_prange(mol, 0, mol.nbas, blksize):
q0, q1 = q1, q1 + nf
dm2_ao = lib.einsum('ijw,pi,qj->pqw', dm2buf, mo_cas[p0:p1],
mo_cas[q0:q1])
shls_slice = (shl0, shl1, b0, b1, 0, mol.nbas, 0, mol.nbas)
eri1 = mol.intor('int2e_ip1',
comp=3,
aosym='s2kl',
shls_slice=shls_slice).reshape(
3, p1 - p0, nf, nao_pair)
de[k] -= numpy.einsum('xijw,ijw->x', eri1, dm2_ao) * 2
eri1 = None
de[k] += numpy.einsum('xij,ij->x', vhf1c[:, p0:p1], dm1[p0:p1]) * 2
de[k] += numpy.einsum('xij,ij->x', vhf1a[:, p0:p1], dm_core[p0:p1]) * 2
dm2 = numpy.zeros((nmo, nmo, nmo, nmo))
for i in range(ncore):
for j in range(ncore):
dm2[i, i, j, j] += 4
dm2[i, j, j, i] -= 2
dm2[i, i, ncore:nocc, ncore:nocc] = casdm1 * 2
dm2[ncore:nocc, ncore:nocc, i, i] = casdm1 * 2
dm2[i, ncore:nocc, ncore:nocc, i] = -casdm1
dm2[ncore:nocc, i, i, ncore:nocc] = -casdm1
dm2[ncore:nocc, ncore:nocc, ncore:nocc, ncore:nocc] = casdm2
eri0 = ao2mo.restore(1, ao2mo.full(mc._scf._eri, mo_coeff), nmo)
Imat = numpy.einsum('pjkl,qjkl->pq', eri0, dm2)
dm1 = numpy.zeros((nmo, nmo))
for i in range(ncore):
dm1[i, i] = 2
dm1[ncore:nocc, ncore:nocc] = casdm1
neleca, nelecb = mol.nelec
h1 = -(mol.intor('int1e_ipkin', comp=3) + mol.intor('int1e_ipnuc', comp=3))
s1 = -mol.intor('int1e_ipovlp', comp=3)
eri1 = mol.intor('int2e_ip1', comp=3).reshape(3, nao, nao, nao, nao)
eri1 = numpy.einsum('xipkl,pj->xijkl', eri1, mo_coeff)
eri1 = numpy.einsum('xijpl,pk->xijkl', eri1, mo_coeff)
eri1 = numpy.einsum('xijkp,pl->xijkl', eri1, mo_coeff)
h0 = reduce(numpy.dot, (mo_coeff.T, mc._scf.get_hcore(), mo_coeff))
g0 = ao2mo.restore(1, ao2mo.full(mol, mo_coeff), nmo)
def hess():
nocc = mol.nelectron // 2
nvir = nmo - nocc
eri_mo = g0
eai = lib.direct_sum('a-i->ai', mo_energy[nocc:], mo_energy[:nocc])
h = eri_mo[nocc:, :nocc, nocc:, :nocc] * 4
h -= numpy.einsum('cdlk->ckdl', eri_mo[nocc:, nocc:, :nocc, :nocc])
h -= numpy.einsum('cldk->ckdl', eri_mo[nocc:, :nocc, nocc:, :nocc])
for a in range(nvir):
for i in range(nocc):
h[a, i, a, i] += eai[a, i]
return -h.reshape(nocc * nvir, -1)
hh = hess()
ee = mo_energy[:, None] - mo_energy
for k, (sh0, sh1, p0, p1) in enumerate(mol.offset_nr_by_atom()):
mol.set_rinv_origin(mol.atom_coord(k))
vrinv = -mol.atom_charge(k) * mol.intor('int1e_iprinv', comp=3)
# 2e AO integrals dot 2pdm
for i in range(3):
g1 = numpy.einsum('pjkl,pi->ijkl', eri1[i, p0:p1], mo_coeff[p0:p1])
g1 = g1 + g1.transpose(1, 0, 2, 3)
g1 = g1 + g1.transpose(2, 3, 0, 1)
g1 *= -1
hx = (numpy.einsum('pq,pi,qj->ij', h1[i, p0:p1], mo_coeff[p0:p1],
mo_coeff) +
reduce(numpy.dot, (mo_coeff.T, vrinv[i], mo_coeff)))
hx = hx + hx.T
sx = numpy.einsum('pq,pi,qj->ij', s1[i, p0:p1], mo_coeff[p0:p1],
mo_coeff)
sx = sx + sx.T
fij = (hx[:neleca, :neleca] - numpy.einsum(
'ij,j->ij', sx[:neleca, :neleca], mo_energy[:neleca]) -
numpy.einsum('kl,ijlk->ij', sx[:neleca, :neleca],
g0[:neleca, :neleca, :neleca, :neleca]) * 2 +
numpy.einsum('kl,iklj->ij', sx[:neleca, :neleca],
g0[:neleca, :neleca, :neleca, :neleca]) +
numpy.einsum('ijkk->ij',
g1[:neleca, :neleca, :neleca, :neleca]) * 2 -
numpy.einsum('ikkj->ij',
g1[:neleca, :neleca, :neleca, :neleca]))
fab = (hx[neleca:, neleca:] - numpy.einsum(
'ij,j->ij', sx[neleca:, neleca:], mo_energy[neleca:]) -
numpy.einsum('kl,ijlk->ij', sx[:neleca, :neleca],
g0[neleca:, neleca:, :neleca, :neleca]) * 2 +
numpy.einsum('kl,iklj->ij', sx[:neleca, :neleca],
g0[neleca:, :neleca, :neleca, neleca:]) +
numpy.einsum('ijkk->ij',
g1[neleca:, neleca:, :neleca, :neleca]) * 2 -
numpy.einsum('ikkj->ij', g1[neleca:, :neleca, :neleca,
neleca:]))
fai = (hx[neleca:, :neleca] - numpy.einsum(
'ai,i->ai', sx[neleca:, :neleca], mo_energy[:neleca]) -
numpy.einsum('kl,ijlk->ij', sx[:neleca, :neleca],
g0[neleca:, :neleca, :neleca, :neleca]) * 2 +
numpy.einsum('kl,iklj->ij', sx[:neleca, :neleca],
g0[neleca:, :neleca, :neleca, :neleca]) +
numpy.einsum('ijkk->ij',
g1[neleca:, :neleca, :neleca, :neleca]) * 2 -
numpy.einsum('ikkj->ij',
g1[neleca:, :neleca, :neleca, :neleca]))
c1 = numpy.zeros((nmo, nmo))
c1[:neleca, :neleca] = -.5 * sx[:neleca, :neleca]
c1[neleca:, neleca:] = -.5 * sx[neleca:, neleca:]
cvo1 = numpy.linalg.solve(hh, fai.ravel()).reshape(-1, neleca)
cov1 = -(sx[neleca:, :neleca] + cvo1).T
c1[neleca:, :neleca] = cvo1
c1[:neleca, neleca:] = cov1
v1 = numpy.einsum('pqai,ai->pq', g0[:, :, neleca:, :neleca],
cvo1) * 4
v1 -= numpy.einsum('paiq,ai->pq', g0[:, neleca:, :neleca, :], cvo1)
v1 -= numpy.einsum('piaq,ai->pq', g0[:, :neleca, neleca:, :], cvo1)
fij += v1[:neleca, :neleca]
fab += v1[neleca:, neleca:]
c1[:ncore,
ncore:neleca] = -fij[:ncore, ncore:] / ee[:ncore, ncore:neleca]
c1[ncore:neleca, :ncore] = -fij[ncore:, :ncore] / ee[
ncore:neleca, :ncore]
m = nocc - neleca
c1[nocc:, neleca:nocc] = -fab[m:, :m] / ee[nocc:, neleca:nocc]
c1[neleca:nocc, nocc:] = -fab[:m, m:] / ee[neleca:nocc, nocc:]
h0c1 = h0.dot(c1)
h0c1 = h0c1 + h0c1.T
g0c1 = numpy.einsum('pjkl,pi->ijkl', g0, c1)
g0c1 = g0c1 + g0c1.transpose(1, 0, 2, 3)
g0c1 = g0c1 + g0c1.transpose(2, 3, 0, 1)
de[k, i] += numpy.einsum('ij,ji', h0c1, dm1)
de[k, i] += numpy.einsum('ijkl,jilk', g0c1, dm2) * .5
de += rhf_grad.grad_nuc(mol)
return de
def kernel(mc, mo_coeff=None, ci=None, atmlst=None, mf_grad=None, verbose=None):
if mo_coeff is None: mo_coeff = mc._scf.mo_coeff
if ci is None: ci = mc.ci
if mf_grad is None: mf_grad = mc._scf.nuc_grad_method()
assert(isinstance(ci, numpy.ndarray))
mol = mc.mol
ncore = mc.ncore
ncas = mc.ncas
nocc = ncore + ncas
nelecas = mc.nelecas
nao, nmo = mo_coeff.shape
nao_pair = nao * (nao+1) // 2
mo_energy = mc._scf.mo_energy
mo_occ = mo_coeff[:,:nocc]
mo_core = mo_coeff[:,:ncore]
mo_cas = mo_coeff[:,ncore:nocc]
neleca, nelecb = mol.nelec
assert(neleca == nelecb)
orbo = mo_coeff[:,:neleca]
orbv = mo_coeff[:,neleca:]
casdm1, casdm2 = mc.fcisolver.make_rdm12(ci, ncas, nelecas)
dm_core = numpy.dot(mo_core, mo_core.T) * 2
dm_cas = reduce(numpy.dot, (mo_cas, casdm1, mo_cas.T))
aapa = ao2mo.kernel(mol, (mo_cas, mo_cas, mo_coeff, mo_cas), compact=False)
aapa = aapa.reshape(ncas,ncas,nmo,ncas)
vj, vk = mc._scf.get_jk(mol, (dm_core, dm_cas))
h1 = mc.get_hcore()
vhf_c = vj[0] - vk[0] * .5
vhf_a = vj[1] - vk[1] * .5
# Imat = h1_{pi} gamma1_{iq} + h2_{pijk} gamma_{iqkj}
Imat = numpy.zeros((nmo,nmo))
Imat[:,:nocc] = reduce(numpy.dot, (mo_coeff.T, h1 + vhf_c + vhf_a, mo_occ)) * 2
Imat[:,ncore:nocc] = reduce(numpy.dot, (mo_coeff.T, h1 + vhf_c, mo_cas, casdm1))
Imat[:,ncore:nocc] += lib.einsum('uviw,vuwt->it', aapa, casdm2)
aapa = vj = vk = vhf_c = vhf_a = h1 = None
ee = mo_energy[:,None] - mo_energy
zvec = numpy.zeros_like(Imat)
zvec[:ncore,ncore:neleca] = Imat[:ncore,ncore:neleca] / -ee[:ncore,ncore:neleca]
zvec[ncore:neleca,:ncore] = Imat[ncore:neleca,:ncore] / -ee[ncore:neleca,:ncore]
zvec[nocc:,neleca:nocc] = Imat[nocc:,neleca:nocc] / -ee[nocc:,neleca:nocc]
zvec[neleca:nocc,nocc:] = Imat[neleca:nocc,nocc:] / -ee[neleca:nocc,nocc:]
zvec_ao = reduce(numpy.dot, (mo_coeff, zvec+zvec.T, mo_coeff.T))
vhf = mc._scf.get_veff(mol, zvec_ao) * 2
xvo = reduce(numpy.dot, (orbv.T, vhf, orbo))
xvo += Imat[neleca:,:neleca] - Imat[:neleca,neleca:].T
def fvind(x):
x = x.reshape(xvo.shape)
dm = reduce(numpy.dot, (orbv, x, orbo.T))
v = mc._scf.get_veff(mol, dm + dm.T)
v = reduce(numpy.dot, (orbv.T, v, orbo))
return v * 2
dm1resp = cphf.solve(fvind, mo_energy, mc._scf.mo_occ, xvo, max_cycle=30)[0]
zvec[neleca:,:neleca] = dm1resp
zeta = numpy.einsum('ij,j->ij', zvec, mo_energy)
zeta = reduce(numpy.dot, (mo_coeff, zeta, mo_coeff.T))
zvec_ao = reduce(numpy.dot, (mo_coeff, zvec+zvec.T, mo_coeff.T))
p1 = numpy.dot(mo_coeff[:,:neleca], mo_coeff[:,:neleca].T)
vhf_s1occ = reduce(numpy.dot, (p1, mc._scf.get_veff(mol, zvec_ao), p1))
Imat[:ncore,ncore:neleca] = 0
Imat[ncore:neleca,:ncore] = 0
Imat[nocc:,neleca:nocc] = 0
Imat[neleca:nocc,nocc:] = 0
Imat[neleca:,:neleca] = Imat[:neleca,neleca:].T
im1 = reduce(numpy.dot, (mo_coeff, Imat, mo_coeff.T))
casci_dm1 = dm_core + dm_cas
hf_dm1 = mc._scf.make_rdm1(mo_coeff, mc._scf.mo_occ)
hcore_deriv = mf_grad.hcore_generator(mol)
s1 = mf_grad.get_ovlp(mol)
diag_idx = numpy.arange(nao)
diag_idx = diag_idx * (diag_idx+1) // 2 + diag_idx
casdm2_cc = casdm2 + casdm2.transpose(0,1,3,2)
dm2buf = ao2mo._ao2mo.nr_e2(casdm2_cc.reshape(ncas**2,ncas**2), mo_cas.T,
(0, nao, 0, nao)).reshape(ncas**2,nao,nao)
dm2buf = lib.pack_tril(dm2buf)
dm2buf[:,diag_idx] *= .5
dm2buf = dm2buf.reshape(ncas,ncas,nao_pair)
casdm2 = casdm2_cc = None
if atmlst is None:
atmlst = range(mol.natm)
aoslices = mol.aoslice_by_atom()
de = numpy.zeros((len(atmlst),3))
max_memory = mc.max_memory - lib.current_memory()[0]
blksize = int(max_memory*.9e6/8 / ((aoslices[:,3]-aoslices[:,2]).max()*nao_pair))
blksize = min(nao, max(2, blksize))
for k, ia in enumerate(atmlst):
shl0, shl1, p0, p1 = aoslices[ia]
h1ao = hcore_deriv(ia)
de[k] += numpy.einsum('xij,ij->x', h1ao, casci_dm1)
de[k] += numpy.einsum('xij,ij->x', h1ao, zvec_ao)
vhf1 = numpy.zeros((3,nao,nao))
q1 = 0
for b0, b1, nf in _shell_prange(mol, 0, mol.nbas, blksize):
q0, q1 = q1, q1 + nf
dm2_ao = lib.einsum('ijw,pi,qj->pqw', dm2buf, mo_cas[p0:p1], mo_cas[q0:q1])
shls_slice = (shl0,shl1,b0,b1,0,mol.nbas,0,mol.nbas)
eri1 = mol.intor('int2e_ip1', comp=3, aosym='s2kl',
shls_slice=shls_slice).reshape(3,p1-p0,nf,nao_pair)
de[k] -= numpy.einsum('xijw,ijw->x', eri1, dm2_ao) * 2
for i in range(3):
eri1tmp = lib.unpack_tril(eri1[i].reshape((p1-p0)*nf,-1))
eri1tmp = eri1tmp.reshape(p1-p0,nf,nao,nao)
de[k,i] -= numpy.einsum('ijkl,ij,kl', eri1tmp, hf_dm1[p0:p1,q0:q1], zvec_ao) * 2
de[k,i] -= numpy.einsum('ijkl,kl,ij', eri1tmp, hf_dm1, zvec_ao[p0:p1,q0:q1]) * 2
de[k,i] += numpy.einsum('ijkl,il,kj', eri1tmp, hf_dm1[p0:p1], zvec_ao[q0:q1])
de[k,i] += numpy.einsum('ijkl,jk,il', eri1tmp, hf_dm1[q0:q1], zvec_ao[p0:p1])
#:vhf1c, vhf1a = mf_grad.get_veff(mol, (dm_core, dm_cas))
#:de[k] += numpy.einsum('xij,ij->x', vhf1c[:,p0:p1], casci_dm1[p0:p1]) * 2
#:de[k] += numpy.einsum('xij,ij->x', vhf1a[:,p0:p1], dm_core[p0:p1]) * 2
de[k,i] -= numpy.einsum('ijkl,lk,ij', eri1tmp, dm_core[q0:q1], casci_dm1[p0:p1]) * 2
de[k,i] += numpy.einsum('ijkl,jk,il', eri1tmp, dm_core[q0:q1], casci_dm1[p0:p1])
de[k,i] -= numpy.einsum('ijkl,lk,ij', eri1tmp, dm_cas[q0:q1], dm_core[p0:p1]) * 2
de[k,i] += numpy.einsum('ijkl,jk,il', eri1tmp, dm_cas[q0:q1], dm_core[p0:p1])
eri1 = eri1tmp = None
de[k] -= numpy.einsum('xij,ij->x', s1[:,p0:p1], im1[p0:p1])
de[k] -= numpy.einsum('xij,ji->x', s1[:,p0:p1], im1[:,p0:p1])
de[k] -= numpy.einsum('xij,ij->x', s1[:,p0:p1], zeta[p0:p1]) * 2
de[k] -= numpy.einsum('xij,ji->x', s1[:,p0:p1], zeta[:,p0:p1]) * 2
de[k] -= numpy.einsum('xij,ij->x', s1[:,p0:p1], vhf_s1occ[p0:p1]) * 2
de[k] -= numpy.einsum('xij,ji->x', s1[:,p0:p1], vhf_s1occ[:,p0:p1]) * 2
de += mf_grad.grad_nuc(mol, atmlst)
return de
def kernel(mp, t2, atmlst=None, mf_grad=None, verbose=logger.INFO):
if mf_grad is None: mf_grad = mp._scf.nuc_grad_method()
log = logger.new_logger(mp, verbose)
time0 = time.clock(), time.time()
log.debug('Build ump2 rdm1 intermediates')
d1 = ump2._gamma1_intermediates(mp, t2)
time1 = log.timer_debug1('rdm1 intermediates', *time0)
log.debug('Build ump2 rdm2 intermediates')
mol = mp.mol
with_frozen = not (mp.frozen is None or mp.frozen is 0)
moidx = mp.get_frozen_mask()
OA_a, VA_a, OF_a, VF_a = mp2_grad._index_frozen_active(moidx[0], mp.mo_occ[0])
OA_b, VA_b, OF_b, VF_b = mp2_grad._index_frozen_active(moidx[1], mp.mo_occ[1])
orboa = mp.mo_coeff[0][:,OA_a]
orbva = mp.mo_coeff[0][:,VA_a]
orbob = mp.mo_coeff[1][:,OA_b]
orbvb = mp.mo_coeff[1][:,VA_b]
nao, nocca = orboa.shape
nvira = orbva.shape[1]
noccb = orbob.shape[1]
nvirb = orbvb.shape[1]
# Partially transform MP2 density matrix and hold it in memory
# The rest transformation are applied during the contraction to ERI integrals
t2aa, t2ab, t2bb = t2
part_dm2aa = _ao2mo.nr_e2(t2aa.reshape(nocca**2,nvira**2),
numpy.asarray(orbva.T, order='F'), (0,nao,0,nao),
's1', 's1').reshape(nocca,nocca,nao,nao)
part_dm2bb = _ao2mo.nr_e2(t2bb.reshape(noccb**2,nvirb**2),
numpy.asarray(orbvb.T, order='F'), (0,nao,0,nao),
's1', 's1').reshape(noccb,noccb,nao,nao)
part_dm2ab = lib.einsum('ijab,pa,qb->ipqj', t2ab, orbva, orbvb)
part_dm2aa = (part_dm2aa.transpose(0,2,3,1) -
part_dm2aa.transpose(0,3,2,1)) * .5
part_dm2bb = (part_dm2bb.transpose(0,2,3,1) -
part_dm2bb.transpose(0,3,2,1)) * .5
hf_dm1a, hf_dm1b = mp._scf.make_rdm1(mp.mo_coeff, mp.mo_occ)
hf_dm1 = hf_dm1a + hf_dm1b
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))
Imata = numpy.zeros((nao,nao))
Imatb = numpy.zeros((nao,nao))
fdm2 = lib.H5TmpFile()
vhf1 = fdm2.create_dataset('vhf1', (len(atmlst),2,3,nao,nao), 'f8')
# 2e AO integrals dot 2pdm
max_memory = max(0, mp.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((2,3,nao,nao))
for b0, b1, nf in mp2_grad._shell_prange(mol, shl0, shl1, blksize):
ip0, ip1 = ip1, ip1 + nf
dm2bufa = lib.einsum('pi,iqrj->pqrj', orboa[ip0:ip1], part_dm2aa)
dm2bufa+= lib.einsum('qi,iprj->pqrj', orboa, part_dm2aa[:,ip0:ip1])
dm2bufa = lib.einsum('pqrj,sj->pqrs', dm2bufa, orboa)
tmp = lib.einsum('pi,iqrj->pqrj', orboa[ip0:ip1], part_dm2ab)
tmp+= lib.einsum('qi,iprj->pqrj', orboa, part_dm2ab[:,ip0:ip1])
dm2bufa+= lib.einsum('pqrj,sj->pqrs', tmp, orbob)
tmp = None
dm2bufa = dm2bufa + dm2bufa.transpose(0,1,3,2)
dm2bufa = lib.pack_tril(dm2bufa.reshape(-1,nao,nao)).reshape(nf,nao,-1)
dm2bufa[:,:,diagidx] *= .5
dm2bufb = lib.einsum('pi,iqrj->pqrj', orbob[ip0:ip1], part_dm2bb)
dm2bufb+= lib.einsum('qi,iprj->pqrj', orbob, part_dm2bb[:,ip0:ip1])
dm2bufb = lib.einsum('pqrj,sj->pqrs', dm2bufb, orbob)
tmp = lib.einsum('iqrj,sj->iqrs', part_dm2ab, orbob[ip0:ip1])
tmp+= lib.einsum('iqrj,sj->iqsr', part_dm2ab[:,:,ip0:ip1], orbob)
dm2bufb+= lib.einsum('pi,iqrs->srpq', orboa, tmp)
tmp = None
dm2bufb = dm2bufb + dm2bufb.transpose(0,1,3,2)
dm2bufb = lib.pack_tril(dm2bufb.reshape(-1,nao,nao)).reshape(nf,nao,-1)
dm2bufb[:,:,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)
Imata += lib.einsum('ipx,iqx->pq', eri0.reshape(nf,nao,-1), dm2bufa)
Imatb += lib.einsum('ipx,iqx->pq', eri0.reshape(nf,nao,-1), dm2bufb)
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, dm2bufa) * 2
de[k] -= numpy.einsum('xijk,ijk->x', eri1, dm2bufb) * 2
dm2bufa = dm2bufb = 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[0,i] -= numpy.einsum('ijkl,il->kj', eri1tmp, hf_dm1a[ip0:ip1])
vhf[1,i] -= numpy.einsum('ijkl,il->kj', eri1tmp, hf_dm1b[ip0:ip1])
vhf[:,i,ip0:ip1] += numpy.einsum('ijkl,kl->ij', eri1tmp, hf_dm1)
vhf[0,i,ip0:ip1] -= numpy.einsum('ijkl,jk->il', eri1tmp, hf_dm1a)
vhf[1,i,ip0:ip1] -= numpy.einsum('ijkl,jk->il', eri1tmp, hf_dm1b)
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)
# Recompute nocc, nvir to include the frozen orbitals and make contraction for
# the 1-particle quantities, see also the kernel function in uccsd_grad module.
mo_a, mo_b = mp.mo_coeff
mo_ea, mo_eb = mp._scf.mo_energy
nao, nmoa = mo_a.shape
nmob = mo_b.shape[1]
nocca = numpy.count_nonzero(mp.mo_occ[0] > 0)
noccb = numpy.count_nonzero(mp.mo_occ[1] > 0)
s0 = mp._scf.get_ovlp()
Imata = reduce(numpy.dot, (mo_a.T, Imata, s0, mo_a)) * -1
Imatb = reduce(numpy.dot, (mo_b.T, Imatb, s0, mo_b)) * -1
dm1a = numpy.zeros((nmoa,nmoa))
dm1b = numpy.zeros((nmob,nmob))
doo, dOO = d1[0]
dvv, dVV = d1[1]
if with_frozen:
dco = Imata[OF_a[:,None],OA_a] / (mo_ea[OF_a,None] - mo_ea[OA_a])
dfv = Imata[VF_a[:,None],VA_a] / (mo_ea[VF_a,None] - mo_ea[VA_a])
dm1a[OA_a[:,None],OA_a] = (doo + doo.T) * .5
dm1a[OF_a[:,None],OA_a] = dco
dm1a[OA_a[:,None],OF_a] = dco.T
dm1a[VA_a[:,None],VA_a] = (dvv + dvv.T) * .5
dm1a[VF_a[:,None],VA_a] = dfv
dm1a[VA_a[:,None],VF_a] = dfv.T
dco = Imatb[OF_b[:,None],OA_b] / (mo_eb[OF_b,None] - mo_eb[OA_b])
dfv = Imatb[VF_b[:,None],VA_b] / (mo_eb[VF_b,None] - mo_eb[VA_b])
dm1b[OA_b[:,None],OA_b] = (dOO + dOO.T) * .5
dm1b[OF_b[:,None],OA_b] = dco
dm1b[OA_b[:,None],OF_b] = dco.T
dm1b[VA_b[:,None],VA_b] = (dVV + dVV.T) * .5
dm1b[VF_b[:,None],VA_b] = dfv
dm1b[VA_b[:,None],VF_b] = dfv.T
else:
dm1a[:nocca,:nocca] = (doo + doo.T) * .5
dm1a[nocca:,nocca:] = (dvv + dvv.T) * .5
dm1b[:noccb,:noccb] = (dOO + dOO.T) * .5
dm1b[noccb:,noccb:] = (dVV + dVV.T) * .5
dm1 = (reduce(numpy.dot, (mo_a, dm1a, mo_a.T)),
reduce(numpy.dot, (mo_b, dm1b, mo_b.T)))
vhf = mp._scf.get_veff(mp.mol, dm1)
Xvo = reduce(numpy.dot, (mo_a[:,nocca:].T, vhf[0], mo_a[:,:nocca]))
XVO = reduce(numpy.dot, (mo_b[:,noccb:].T, vhf[1], mo_b[:,:noccb]))
Xvo+= Imata[:nocca,nocca:].T - Imata[nocca:,:nocca]
XVO+= Imatb[:noccb,noccb:].T - Imatb[noccb:,:noccb]
dm1_resp = _response_dm1(mp, (Xvo,XVO))
dm1a += dm1_resp[0]
dm1b += dm1_resp[1]
time1 = log.timer_debug1('response_rdm1 intermediates', *time1)
Imata[nocca:,:nocca] = Imata[:nocca,nocca:].T
Imatb[noccb:,:noccb] = Imatb[:noccb,noccb:].T
im1 = reduce(numpy.dot, (mo_a, Imata, mo_a.T))
im1+= reduce(numpy.dot, (mo_b, Imatb, mo_b.T))
time1 = log.timer_debug1('response_rdm1', *time1)
log.debug('h1 and JK1')
hcore_deriv = mf_grad.hcore_generator(mol)
s1 = mf_grad.get_ovlp(mol)
zeta = (mo_ea[:,None] + mo_ea) * .5
zeta[nocca:,:nocca] = mo_ea[:nocca]
zeta[:nocca,nocca:] = mo_ea[:nocca].reshape(-1,1)
zeta_a = reduce(numpy.dot, (mo_a, zeta*dm1a, mo_a.T))
zeta = (mo_eb[:,None] + mo_eb) * .5
zeta[noccb:,:noccb] = mo_eb[:noccb]
zeta[:noccb,noccb:] = mo_eb[:noccb].reshape(-1,1)
zeta_b = reduce(numpy.dot, (mo_b, zeta*dm1b, mo_b.T))
dm1 = (reduce(numpy.dot, (mo_a, dm1a, mo_a.T)),
reduce(numpy.dot, (mo_b, dm1b, mo_b.T)))
vhf_s1occ = mp._scf.get_veff(mol, (dm1[0]+dm1[0].T, dm1[1]+dm1[1].T))
p1a = numpy.dot(mo_a[:,:nocca], mo_a[:,:nocca].T)
p1b = numpy.dot(mo_b[:,:noccb], mo_b[:,:noccb].T)
vhf_s1occ = (reduce(numpy.dot, (p1a, vhf_s1occ[0], p1a)) +
reduce(numpy.dot, (p1b, vhf_s1occ[1], p1b))) * .5
time1 = log.timer_debug1('h1 and JK1', *time1)
# Hartree-Fock part contribution
dm1pa = hf_dm1a + dm1[0]*2
dm1pb = hf_dm1b + dm1[1]*2
dm1 = dm1[0] + dm1[1] + hf_dm1
zeta_a += rhf_grad.make_rdm1e(mo_ea, mo_a, mp.mo_occ[0])
zeta_b += rhf_grad.make_rdm1e(mo_eb, mo_b, mp.mo_occ[1])
zeta = zeta_a + zeta_b
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,0], dm1pa)
de[k] -= numpy.einsum('xij,ij->x', vhf1[k,1], dm1pb)
de += mf_grad.grad_nuc(mol)
log.timer('%s gradients' % mp.__class__.__name__, *time0)
return de
def Lci_dot_dgci_dx(Lci,
weights,
mc,
mo_coeff=None,
ci=None,
atmlst=None,
mf_grad=None,
eris=None,
verbose=None):
''' Modification of pyscf.grad.casscf.kernel to compute instead the CI
Lagrange term nuclear gradient (sum_IJ Lci_IJ d2_Ecas/d_lambda d_PIJ)
This involves removing all core-core and nuclear-nuclear terms and making the substitution
sum_I w_I<L_I|p'q|I> + c.c. -> <0|p'q|0>
sum_I w_I<L_I|p'r'sq|I> + c.c. -> <0|p'r'sq|0>
The active-core terms (sum_I w_I<L_I|x'iyi|I>, sum_I w_I <L_I|x'iiy|I>, c.c.) must be retained.'''
if mo_coeff is None: mo_coeff = mc.mo_coeff
if ci is None: ci = mc.ci
if mf_grad is None: mf_grad = mc._scf.nuc_grad_method()
if mc.frozen is not None:
raise NotImplementedError
t0 = (logger.process_clock(), logger.perf_counter())
mol = mc.mol
ncore = mc.ncore
ncas = mc.ncas
nocc = ncore + ncas
nelecas = mc.nelecas
nao, nmo = mo_coeff.shape
nao_pair = nao * (nao + 1) // 2
mo_occ = mo_coeff[:, :nocc]
mo_core = mo_coeff[:, :ncore]
mo_cas = mo_coeff[:, ncore:nocc]
# MRH: TDMs + c.c. instead of RDMs; 06/30/2020: new interface in mcscf.addons makes this much more transparent
casdm1, casdm2 = mc.fcisolver.trans_rdm12(Lci, ci, ncas, nelecas)
casdm1 += casdm1.transpose(1, 0)
casdm2 += casdm2.transpose(1, 0, 3, 2)
# gfock = Generalized Fock, Adv. Chem. Phys., 69, 63
dm_core = np.dot(mo_core, mo_core.T) * 2
dm_cas = reduce(np.dot, (mo_cas, casdm1, mo_cas.T))
aapa = np.zeros((ncas, ncas, nmo, ncas), dtype=dm_cas.dtype)
for i in range(nmo):
aapa[:, :, i, :] = eris.ppaa[i][ncore:nocc, :, :].transpose(1, 2, 0)
vj, vk = mc._scf.get_jk(mol, (dm_core, dm_cas))
h1 = mc.get_hcore()
vhf_c = vj[0] - vk[0] * .5
vhf_a = vj[1] - vk[1] * .5
# MRH: delete h1 + vhf_c from the first line below (core and core-core stuff)
# Also extend gfock to span the whole space
gfock = np.zeros_like(dm_cas)
gfock[:, :nocc] = reduce(np.dot, (mo_coeff.T, vhf_a, mo_occ)) * 2
gfock[:, ncore:nocc] = reduce(np.dot,
(mo_coeff.T, h1 + vhf_c, mo_cas, casdm1))
gfock[:, ncore:nocc] += np.einsum('uvpw,vuwt->pt', aapa, casdm2)
dme0 = reduce(np.dot, (mo_coeff, (gfock + gfock.T) * .5, mo_coeff.T))
aapa = vj = vk = vhf_c = vhf_a = h1 = gfock = None
vj, vk = mf_grad.get_jk(mol, (dm_core, dm_cas))
vhf1c, vhf1a = vj - vk * 0.5
#vhf1c, vhf1a = mf_grad.get_veff(mol, (dm_core, dm_cas))
hcore_deriv = mf_grad.hcore_generator(mol)
s1 = mf_grad.get_ovlp(mol)
diag_idx = np.arange(nao)
diag_idx = diag_idx * (diag_idx + 1) // 2 + diag_idx
casdm2_cc = casdm2 + casdm2.transpose(0, 1, 3, 2)
dm2buf = ao2mo._ao2mo.nr_e2(casdm2_cc.reshape(ncas**2, ncas**2), mo_cas.T,
(0, nao, 0, nao)).reshape(ncas**2, nao, nao)
dm2buf = lib.pack_tril(dm2buf)
dm2buf[:, diag_idx] *= .5
dm2buf = dm2buf.reshape(ncas, ncas, nao_pair)
casdm2 = casdm2_cc = None
if atmlst is None:
atmlst = range(mol.natm)
aoslices = mol.aoslice_by_atom()
de_hcore = np.zeros((len(atmlst), 3))
de_renorm = np.zeros((len(atmlst), 3))
de_eri = np.zeros((len(atmlst), 3))
de = np.zeros((len(atmlst), 3))
max_memory = mc.max_memory - lib.current_memory()[0]
blksize = int(max_memory * .9e6 / 8 /
(4 * (aoslices[:, 3] - aoslices[:, 2]).max() * nao_pair))
# MRH: 3 components of eri array and 1 density matrix array: FOUR arrays of this size are required!
blksize = min(nao, max(2, blksize))
logger.info(
mc,
'SA-CASSCF Lci_dot_dgci memory remaining for eri manipulation: {} MB; using blocksize = {}'
.format(max_memory, blksize))
t0 = logger.timer(mc, 'SA-CASSCF Lci_dot_dgci 1-electron part', *t0)
for k, ia in enumerate(atmlst):
shl0, shl1, p0, p1 = aoslices[ia]
h1ao = hcore_deriv(ia)
# MRH: dm1 -> dm_cas in the line below
de_hcore[k] += np.einsum('xij,ij->x', h1ao, dm_cas)
de_renorm[k] -= np.einsum('xij,ij->x', s1[:, p0:p1], dme0[p0:p1]) * 2
q1 = 0
for b0, b1, nf in _shell_prange(mol, 0, mol.nbas, blksize):
q0, q1 = q1, q1 + nf
dm2_ao = lib.einsum('ijw,pi,qj->pqw', dm2buf, mo_cas[p0:p1],
mo_cas[q0:q1])
shls_slice = (shl0, shl1, b0, b1, 0, mol.nbas, 0, mol.nbas)
gc.collect()
eri1 = mol.intor('int2e_ip1',
comp=3,
aosym='s2kl',
shls_slice=shls_slice).reshape(
3, p1 - p0, nf, nao_pair)
de_eri[k] -= np.einsum('xijw,ijw->x', eri1, dm2_ao) * 2
eri1 = dm2_ao = None
gc.collect()
t0 = logger.timer(
mc, 'SA-CASSCF Lci_dot_dgci atom {} ({},{}|{})'.format(
ia, p1 - p0, nf, nao_pair), *t0)
# MRH: dm1 -> dm_cas in the line below. Also eliminate core-core terms
de_eri[k] += np.einsum('xij,ij->x', vhf1c[:, p0:p1], dm_cas[p0:p1]) * 2
de_eri[k] += np.einsum('xij,ij->x', vhf1a[:, p0:p1],
dm_core[p0:p1]) * 2
logger.debug(mc, "CI lagrange hcore component:\n{}".format(de_hcore))
logger.debug(mc, "CI lagrange renorm component:\n{}".format(de_renorm))
logger.debug(mc, "CI lagrange eri component:\n{}".format(de_eri))
de = de_hcore + de_renorm + de_eri
return de
def Lorb_dot_dgorb_dx(Lorb,
mc,
mo_coeff=None,
ci=None,
atmlst=None,
mf_grad=None,
eris=None,
verbose=None):
''' Modification of pyscf.grad.casscf.kernel to compute instead the orbital
Lagrange term nuclear gradient (sum_pq Lorb_pq d2_Ecas/d_lambda d_kpq)
This involves removing nuclear-nuclear terms and making the substitution
(D_[p]q + D_p[q]) -> D_pq
(d_[p]qrs + d_pq[r]s + d_p[q]rs + d_pqr[s]) -> d_pqrs
Where [] around an index implies contraction with Lorb from the left, so that the external index
(regardless of whether the index on the rdm is bra or ket) is always the first index of Lorb. '''
# dmo = smoT.dao.smo
# dao = mo.dmo.moT
t0 = (logger.process_clock(), logger.perf_counter())
if mo_coeff is None: mo_coeff = mc.mo_coeff
if ci is None: ci = mc.ci
if mf_grad is None: mf_grad = mc._scf.nuc_grad_method()
if mc.frozen is not None:
raise NotImplementedError
mol = mc.mol
ncore = mc.ncore
ncas = mc.ncas
nocc = ncore + ncas
nelecas = mc.nelecas
nao, nmo = mo_coeff.shape
nao_pair = nao * (nao + 1) // 2
mo_core = mo_coeff[:, :ncore]
mo_cas = mo_coeff[:, ncore:nocc]
# MRH: new 'effective' MO coefficients including contraction from the Lagrange multipliers
moL_coeff = np.dot(mo_coeff, Lorb)
s0_inv = np.dot(mo_coeff, mo_coeff.T)
moL_core = moL_coeff[:, :ncore]
moL_cas = moL_coeff[:, ncore:nocc]
# MRH: these SHOULD be state-averaged! Use the actual sacasscf object!
casdm1, casdm2 = mc.fcisolver.make_rdm12(ci, ncas, nelecas)
# gfock = Generalized Fock, Adv. Chem. Phys., 69, 63
# MRH: each index exactly once!
dm_core = np.dot(mo_core, mo_core.T) * 2
dm_cas = reduce(np.dot, (mo_cas, casdm1, mo_cas.T))
# MRH: new density matrix terms
dmL_core = np.dot(moL_core, mo_core.T) * 2
dmL_cas = reduce(np.dot, (moL_cas, casdm1, mo_cas.T))
dmL_core += dmL_core.T
dmL_cas += dmL_cas.T
dm1 = dm_core + dm_cas
dm1L = dmL_core + dmL_cas
# MRH: end new density matrix terms
# MRH: wrap the integral instead of the density matrix. I THINK the sign is the same!
# mo sets 0 and 2 should be transposed, 1 and 3 should be not transposed; this will lead to correct sign
# Except I can't do this for the external index, because the external index is contracted to ovlp matrix,
# not the 2RDM
aapa = np.zeros((ncas, ncas, nmo, ncas), dtype=dm_cas.dtype)
aapaL = np.zeros((ncas, ncas, nmo, ncas), dtype=dm_cas.dtype)
for i in range(nmo):
jbuf = eris.ppaa[i]
kbuf = eris.papa[i]
aapa[:, :, i, :] = jbuf[ncore:nocc, :, :].transpose(1, 2, 0)
aapaL[:, :, i, :] += np.tensordot(jbuf,
Lorb[:, ncore:nocc],
axes=((0), (0)))
kbuf = np.tensordot(kbuf, Lorb[:, ncore:nocc],
axes=((1), (0))).transpose(1, 2, 0)
aapaL[:, :, i, :] += kbuf + kbuf.transpose(1, 0, 2)
# MRH: new vhf terms
vj, vk = mc._scf.get_jk(mol, (dm_core, dm_cas))
vjL, vkL = mc._scf.get_jk(mol, (dmL_core, dmL_cas))
h1 = mc.get_hcore()
vhf_c = vj[0] - vk[0] * .5
vhf_a = vj[1] - vk[1] * .5
vhfL_c = vjL[0] - vkL[0] * .5
vhfL_a = vjL[1] - vkL[1] * .5
# MRH: I rewrote this Feff calculation completely, double-check it
gfock = np.dot(h1, dm1L) # h1e
gfock += np.dot((vhf_c + vhf_a),
dmL_core) # core-core and active-core, 2nd 1RDM linked
gfock += np.dot((vhfL_c + vhfL_a),
dm_core) # core-core and active-core, 1st 1RDM linked
gfock += np.dot(vhfL_c, dm_cas) # core-active, 1st 1RDM linked
gfock += np.dot(vhf_c, dmL_cas) # core-active, 2nd 1RDM linked
gfock = np.dot(
s0_inv, gfock
) # Definition of quantity is in MO's; going (AO->MO->AO) incurs an inverse ovlp
gfock += reduce(np.dot, (mo_coeff, np.einsum(
'uviw,uvtw->it', aapaL, casdm2), mo_cas.T)) # active-active
# MRH: I have to contract this external 2RDM index explicitly on the 2RDM but fortunately I can do so here
gfock += reduce(
np.dot,
(mo_coeff, np.einsum('uviw,vuwt->it', aapa, casdm2), moL_cas.T))
# MRH: As of 04/18/2019, the two-body part of this is including aapaL is definitely, unambiguously correct
dme0 = (gfock +
gfock.T) / 2 # This transpose is for the overlap matrix later on
aapa = vj = vk = vhf_c = vhf_a = None
vj, vk = mf_grad.get_jk(mol, (dm_core, dm_cas, dmL_core, dmL_cas))
vhf1c, vhf1a, vhf1cL, vhf1aL = vj - vk * 0.5
#vhf1c, vhf1a, vhf1cL, vhf1aL = mf_grad.get_veff(mol, (dm_core, dm_cas, dmL_core, dmL_cas))
hcore_deriv = mf_grad.hcore_generator(mol)
s1 = mf_grad.get_ovlp(mol)
diag_idx = np.arange(nao)
diag_idx = diag_idx * (diag_idx + 1) // 2 + diag_idx
casdm2_cc = casdm2 + casdm2.transpose(0, 1, 3, 2)
dm2buf = ao2mo._ao2mo.nr_e2(casdm2_cc.reshape(ncas**2, ncas**2), mo_cas.T,
(0, nao, 0, nao)).reshape(ncas**2, nao, nao)
# MRH: contract the final two indices of the active-active 2RDM with L as you change to AOs
# note tensordot always puts indices in the order of the arguments.
dm2Lbuf = np.zeros((ncas**2, nmo, nmo))
# MRH: The second line below transposes the L; the third line transposes the derivative later on
# Both the L and the derivative have to explore all indices
dm2Lbuf[:, :, ncore:nocc] = np.tensordot(
Lorb[:, ncore:nocc], casdm2,
axes=(1, 2)).transpose(1, 2, 0, 3).reshape(ncas**2, nmo, ncas)
dm2Lbuf[:, ncore:nocc, :] += np.tensordot(
Lorb[:, ncore:nocc], casdm2,
axes=(1, 3)).transpose(1, 2, 3, 0).reshape(ncas**2, ncas, nmo)
dm2Lbuf += dm2Lbuf.transpose(0, 2, 1)
dm2Lbuf = np.ascontiguousarray(dm2Lbuf)
dm2Lbuf = ao2mo._ao2mo.nr_e2(dm2Lbuf.reshape(ncas**2, nmo**2), mo_coeff.T,
(0, nao, 0, nao)).reshape(ncas**2, nao, nao)
dm2buf = lib.pack_tril(dm2buf)
dm2buf[:, diag_idx] *= .5
dm2buf = dm2buf.reshape(ncas, ncas, nao_pair)
dm2Lbuf = lib.pack_tril(dm2Lbuf)
dm2Lbuf[:, diag_idx] *= .5
dm2Lbuf = dm2Lbuf.reshape(ncas, ncas, nao_pair)
if atmlst is None:
atmlst = list(range(mol.natm))
aoslices = mol.aoslice_by_atom()
de_hcore = np.zeros((len(atmlst), 3))
de_renorm = np.zeros((len(atmlst), 3))
de_eri = np.zeros((len(atmlst), 3))
de = np.zeros((len(atmlst), 3))
max_memory = mc.max_memory - lib.current_memory()[0]
blksize = int(max_memory * .9e6 / 8 /
(4 * (aoslices[:, 3] - aoslices[:, 2]).max() * nao_pair))
# MRH: 3 components of eri array and 1 density matrix array: FOUR arrays of this size are required!
blksize = min(nao, max(2, blksize))
logger.info(
mc,
'SA-CASSCF Lorb_dot_dgorb memory remaining for eri manipulation: {} MB; using blocksize = {}'
.format(max_memory, blksize))
t0 = logger.timer(mc, 'SA-CASSCF Lorb_dot_dgorb 1-electron part', *t0)
for k, ia in enumerate(atmlst):
shl0, shl1, p0, p1 = aoslices[ia]
h1ao = hcore_deriv(ia)
# MRH: h1e and Feff terms
de_hcore[k] += np.einsum('xij,ij->x', h1ao, dm1L)
de_renorm[k] -= np.einsum('xij,ij->x', s1[:, p0:p1], dme0[p0:p1]) * 2
q1 = 0
for b0, b1, nf in _shell_prange(mol, 0, mol.nbas, blksize):
q0, q1 = q1, q1 + nf
dm2_ao = lib.einsum('ijw,pi,qj->pqw', dm2Lbuf, mo_cas[p0:p1],
mo_cas[q0:q1])
# MRH: now contract the first two indices of the active-active 2RDM with L as you go from MOs to AOs
dm2_ao += lib.einsum('ijw,pi,qj->pqw', dm2buf, moL_cas[p0:p1],
mo_cas[q0:q1])
dm2_ao += lib.einsum('ijw,pi,qj->pqw', dm2buf, mo_cas[p0:p1],
moL_cas[q0:q1])
shls_slice = (shl0, shl1, b0, b1, 0, mol.nbas, 0, mol.nbas)
gc.collect()
eri1 = mol.intor('int2e_ip1',
comp=3,
aosym='s2kl',
shls_slice=shls_slice).reshape(
3, p1 - p0, nf, nao_pair)
# MRH: I still don't understand why there is a minus here!
de_eri[k] -= np.einsum('xijw,ijw->x', eri1, dm2_ao) * 2
eri1 = dm2_ao = None
gc.collect()
t0 = logger.timer(
mc, 'SA-CASSCF Lorb_dot_dgorb atom {} ({},{}|{})'.format(
ia, p1 - p0, nf, nao_pair), *t0)
# MRH: core-core and core-active 2RDM terms
de_eri[k] += np.einsum('xij,ij->x', vhf1c[:, p0:p1], dm1L[p0:p1]) * 2
de_eri[k] += np.einsum('xij,ij->x', vhf1cL[:, p0:p1], dm1[p0:p1]) * 2
# MRH: active-core 2RDM terms
de_eri[k] += np.einsum('xij,ij->x', vhf1a[:, p0:p1],
dmL_core[p0:p1]) * 2
de_eri[k] += np.einsum('xij,ij->x', vhf1aL[:, p0:p1],
dm_core[p0:p1]) * 2
# MRH: deleted the nuclear-nuclear part to avoid double-counting
# lesson learned from debugging - mol.intor computes -1 * the derivative and only
# for one index
# on the other hand, mf_grad.hcore_generator computes the actual derivative of
# h1 for both indices and with the correct sign
logger.debug(mc, "Orb lagrange hcore component:\n{}".format(de_hcore))
logger.debug(mc, "Orb lagrange renorm component:\n{}".format(de_renorm))
logger.debug(mc, "Orb lagrange eri component:\n{}".format(de_eri))
de = de_hcore + de_renorm + de_eri
return de
def kernel(mc, mo_coeff=None, ci=None, atmlst=None, mf_grad=None,
verbose=None):
if mo_coeff is None: mo_coeff = mc._scf.mo_coeff
if ci is None: ci = mc.ci
if mf_grad is None: mf_grad = mc._scf.nuc_grad_method()
mol = mc.mol
ncore = mc.ncore
ncas = mc.ncas
nocc = ncore + ncas
nelecas = mc.nelecas
nao, nmo = mo_coeff.shape
nao_pair = nao * (nao+1) // 2
mo_energy = mc._scf.mo_energy
hcore_deriv = mf_grad.hcore_generator(mol)
s1 = mf_grad.get_ovlp(mol)
mo_occ = mo_coeff[:,:nocc]
mo_core = mo_coeff[:,:ncore]
mo_cas = mo_coeff[:,ncore:nocc]
casdm1, casdm2 = mc.fcisolver.make_rdm12(mc.ci, ncas, nelecas)
# gfock = Generalized Fock, Adv. Chem. Phys., 69, 63
dm_core = numpy.dot(mo_core, mo_core.T) * 2
dm_cas = reduce(numpy.dot, (mo_cas, casdm1, mo_cas.T))
aapa = ao2mo.kernel(mol, (mo_cas, mo_cas, mo_occ, mo_cas), compact=False)
aapa = aapa.reshape(ncas,ncas,nocc,ncas)
vj, vk = mc._scf.get_jk(mol, (dm_core, dm_cas))
h1 = mc.get_hcore()
vhf_c = vj[0] - vk[0] * .5
vhf_a = vj[1] - vk[1] * .5
gfock = reduce(numpy.dot, (mo_occ.T, h1 + vhf_c + vhf_a, mo_occ)) * 2
gfock[:,ncore:nocc] = reduce(numpy.dot, (mo_occ.T, h1 + vhf_c, mo_cas, casdm1))
gfock[:,ncore:nocc] += numpy.einsum('uviw,vuwt->it', aapa, casdm2)
dme0 = reduce(numpy.dot, (mo_occ, (gfock+gfock.T)*.5, mo_occ.T))
aapa = vj = vk = vhf_c = vhf_a = h1 = gfock = None
dm1 = dm_core + dm_cas
vhf1c, vhf1a = mf_grad.get_veff(mol, (dm_core, dm_cas))
diag_idx = numpy.arange(nao)
diag_idx = diag_idx * (diag_idx+1) // 2 + diag_idx
casdm2_cc = casdm2 + casdm2.transpose(0,1,3,2)
dm2buf = ao2mo._ao2mo.nr_e2(casdm2_cc.reshape(ncas**2,ncas**2), mo_cas.T,
(0, nao, 0, nao)).reshape(ncas**2,nao,nao)
dm2buf = lib.pack_tril(dm2buf)
dm2buf[:,diag_idx] *= .5
dm2buf = dm2buf.reshape(ncas,ncas,nao_pair)
#casdm2 = casdm2_cc = None
atmlst = range(mol.natm)
aoslices = mol.aoslice_by_atom()
de = numpy.zeros((len(atmlst),3))
max_memory = mc.max_memory - lib.current_memory()[0]
blksize = int(max_memory*.9e6/8 / ((aoslices[:,3]-aoslices[:,2]).max()*nao_pair))
blksize = min(nao, max(2, blksize))
for k, ia in enumerate(atmlst):
shl0, shl1, p0, p1 = aoslices[ia]
h1ao = hcore_deriv(ia)
de[k] += numpy.einsum('xij,ij->x', h1ao, dm1)
#de[k] -= numpy.einsum('xij,ij->x', s1[:,p0:p1], dme0[p0:p1]) * 2
q1 = 0
for b0, b1, nf in _shell_prange(mol, 0, mol.nbas, blksize):
q0, q1 = q1, q1 + nf
dm2_ao = lib.einsum('ijw,pi,qj->pqw', dm2buf, mo_cas[p0:p1], mo_cas[q0:q1])
shls_slice = (shl0,shl1,b0,b1,0,mol.nbas,0,mol.nbas)
eri1 = mol.intor('int2e_ip1', comp=3, aosym='s2kl',
shls_slice=shls_slice).reshape(3,p1-p0,nf,nao_pair)
de[k] -= numpy.einsum('xijw,ijw->x', eri1, dm2_ao) * 2
eri1 = None
de[k] += numpy.einsum('xij,ij->x', vhf1c[:,p0:p1], dm1[p0:p1]) * 2
de[k] += numpy.einsum('xij,ij->x', vhf1a[:,p0:p1], dm_core[p0:p1]) * 2
dm2 = numpy.zeros((nmo,nmo,nmo,nmo))
for i in range(ncore):
for j in range(ncore):
dm2[i,i,j,j] += 4
dm2[i,j,j,i] -= 2
dm2[i,i,ncore:nocc,ncore:nocc] = casdm1 * 2
dm2[ncore:nocc,ncore:nocc,i,i] = casdm1 * 2
dm2[i,ncore:nocc,ncore:nocc,i] =-casdm1
dm2[ncore:nocc,i,i,ncore:nocc] =-casdm1
dm2[ncore:nocc,ncore:nocc,ncore:nocc,ncore:nocc] = casdm2
eri0 = ao2mo.restore(1, ao2mo.full(mc._scf._eri, mo_coeff), nmo)
Imat = numpy.einsum('pjkl,qjkl->pq', eri0, dm2)
dm1 = numpy.zeros((nmo,nmo))
for i in range(ncore):
dm1[i,i] = 2
dm1[ncore:nocc,ncore:nocc] = casdm1
neleca, nelecb = mol.nelec
h1 =-(mol.intor('int1e_ipkin', comp=3)
+mol.intor('int1e_ipnuc', comp=3))
s1 =-mol.intor('int1e_ipovlp', comp=3)
eri1 = mol.intor('int2e_ip1', comp=3).reshape(3,nao,nao,nao,nao)
eri1 = numpy.einsum('xipkl,pj->xijkl', eri1, mo_coeff)
eri1 = numpy.einsum('xijpl,pk->xijkl', eri1, mo_coeff)
eri1 = numpy.einsum('xijkp,pl->xijkl', eri1, mo_coeff)
h0 = reduce(numpy.dot, (mo_coeff.T, mc._scf.get_hcore(), mo_coeff))
g0 = ao2mo.restore(1, ao2mo.full(mol, mo_coeff), nmo)
def hess():
nocc = mol.nelectron//2
nvir = nmo - nocc
eri_mo = g0
eai = lib.direct_sum('a-i->ai', mo_energy[nocc:], mo_energy[:nocc])
h = eri_mo[nocc:,:nocc,nocc:,:nocc] * 4
h-= numpy.einsum('cdlk->ckdl', eri_mo[nocc:,nocc:,:nocc,:nocc])
h-= numpy.einsum('cldk->ckdl', eri_mo[nocc:,:nocc,nocc:,:nocc])
for a in range(nvir):
for i in range(nocc):
h[a,i,a,i] += eai[a,i]
return -h.reshape(nocc*nvir,-1)
hh = hess()
ee = mo_energy[:,None] - mo_energy
for k,(sh0, sh1, p0, p1) in enumerate(mol.offset_nr_by_atom()):
mol.set_rinv_origin(mol.atom_coord(k))
vrinv = -mol.atom_charge(k) * mol.intor('int1e_iprinv', comp=3)
# 2e AO integrals dot 2pdm
for i in range(3):
g1 = numpy.einsum('pjkl,pi->ijkl', eri1[i,p0:p1], mo_coeff[p0:p1])
g1 = g1 + g1.transpose(1,0,2,3)
g1 = g1 + g1.transpose(2,3,0,1)
g1 *= -1
hx =(numpy.einsum('pq,pi,qj->ij', h1[i,p0:p1], mo_coeff[p0:p1], mo_coeff)
+ reduce(numpy.dot, (mo_coeff.T, vrinv[i], mo_coeff)))
hx = hx + hx.T
sx = numpy.einsum('pq,pi,qj->ij', s1[i,p0:p1], mo_coeff[p0:p1], mo_coeff)
sx = sx + sx.T
fij =(hx[:neleca,:neleca]
- numpy.einsum('ij,j->ij', sx[:neleca,:neleca], mo_energy[:neleca])
- numpy.einsum('kl,ijlk->ij', sx[:neleca,:neleca],
g0[:neleca,:neleca,:neleca,:neleca]) * 2
+ numpy.einsum('kl,iklj->ij', sx[:neleca,:neleca],
g0[:neleca,:neleca,:neleca,:neleca])
+ numpy.einsum('ijkk->ij', g1[:neleca,:neleca,:neleca,:neleca]) * 2
- numpy.einsum('ikkj->ij', g1[:neleca,:neleca,:neleca,:neleca]))
fab =(hx[neleca:,neleca:]
- numpy.einsum('ij,j->ij', sx[neleca:,neleca:], mo_energy[neleca:])
- numpy.einsum('kl,ijlk->ij', sx[:neleca,:neleca],
g0[neleca:,neleca:,:neleca,:neleca]) * 2
+ numpy.einsum('kl,iklj->ij', sx[:neleca,:neleca],
g0[neleca:,:neleca,:neleca,neleca:])
+ numpy.einsum('ijkk->ij', g1[neleca:,neleca:,:neleca,:neleca]) * 2
- numpy.einsum('ikkj->ij', g1[neleca:,:neleca,:neleca,neleca:]))
fai =(hx[neleca:,:neleca]
- numpy.einsum('ai,i->ai', sx[neleca:,:neleca], mo_energy[:neleca])
- numpy.einsum('kl,ijlk->ij', sx[:neleca,:neleca],
g0[neleca:,:neleca,:neleca,:neleca]) * 2
+ numpy.einsum('kl,iklj->ij', sx[:neleca,:neleca],
g0[neleca:,:neleca,:neleca,:neleca])
+ numpy.einsum('ijkk->ij', g1[neleca:,:neleca,:neleca,:neleca]) * 2
- numpy.einsum('ikkj->ij', g1[neleca:,:neleca,:neleca,:neleca]))
c1 = numpy.zeros((nmo,nmo))
c1[:neleca,:neleca] = -.5 * sx[:neleca,:neleca]
c1[neleca:,neleca:] = -.5 * sx[neleca:,neleca:]
cvo1 = numpy.linalg.solve(hh, fai.ravel()).reshape(-1,neleca)
cov1 = -(sx[neleca:,:neleca] + cvo1).T
c1[neleca:,:neleca] = cvo1
c1[:neleca,neleca:] = cov1
v1 = numpy.einsum('pqai,ai->pq', g0[:,:,neleca:,:neleca], cvo1) * 4
v1-= numpy.einsum('paiq,ai->pq', g0[:,neleca:,:neleca,:], cvo1)
v1-= numpy.einsum('piaq,ai->pq', g0[:,:neleca,neleca:,:], cvo1)
fij += v1[:neleca,:neleca]
fab += v1[neleca:,neleca:]
c1[:ncore,ncore:neleca] = -fij[:ncore,ncore:] / ee[:ncore,ncore:neleca]
c1[ncore:neleca,:ncore] = -fij[ncore:,:ncore] / ee[ncore:neleca,:ncore]
m = nocc - neleca
c1[nocc:,neleca:nocc] = -fab[m:,:m] / ee[nocc:,neleca:nocc]
c1[neleca:nocc,nocc:] = -fab[:m,m:] / ee[neleca:nocc,nocc:]
h0c1 = h0.dot(c1)
h0c1 = h0c1 + h0c1.T
g0c1 = numpy.einsum('pjkl,pi->ijkl', g0, c1)
g0c1 = g0c1 + g0c1.transpose(1,0,2,3)
g0c1 = g0c1 + g0c1.transpose(2,3,0,1)
de[k,i] += numpy.einsum('ij,ji', h0c1, dm1)
de[k,i] += numpy.einsum('ijkl,jilk', g0c1, dm2)*.5
de += rhf_grad.grad_nuc(mol)
return de
def Lci_dot_dgci_dx(Lci,
weights,
mc,
mo_coeff=None,
ci=None,
atmlst=None,
mf_grad=None,
verbose=None):
''' Modification of pyscf.grad.casscf.kernel to compute instead the CI
Lagrange term nuclear gradient (sum_IJ Lci_IJ d2_Ecas/d_lambda d_PIJ)
This involves removing all core-core and nuclear-nuclear terms and making the substitution
sum_I w_I<L_I|p'q|I> + c.c. -> <0|p'q|0>
sum_I w_I<L_I|p'r'sq|I> + c.c. -> <0|p'r'sq|0>
The active-core terms (sum_I w_I<L_I|x'iyi|I>, sum_I w_I <L_I|x'iiy|I>, c.c.) must be retained.'''
if mo_coeff is None: mo_coeff = mc.mo_coeff
if ci is None: ci = mc.ci
if mf_grad is None: mf_grad = mc._scf.nuc_grad_method()
if mc.frozen is not None:
raise NotImplementedError
mol = mc.mol
ncore = mc.ncore
ncas = mc.ncas
nocc = ncore + ncas
nelecas = mc.nelecas
nao, nmo = mo_coeff.shape
nao_pair = nao * (nao + 1) // 2
nroots = ci.shape[0]
mo_occ = mo_coeff[:, :nocc]
mo_core = mo_coeff[:, :ncore]
mo_cas = mo_coeff[:, ncore:nocc]
# MRH: TDMs + c.c. instead of RDMs
casdm1 = np.zeros((nroots, ncas, ncas))
casdm2 = np.zeros((nroots, ncas, ncas, ncas, ncas))
for iroot in range(nroots):
#print ("norm of Lci, ci for root {}: {} {}".format (iroot, linalg.norm (Lci[iroot]), linalg.norm (ci[iroot])))
casdm1[iroot], casdm2[iroot] = mc.fcisolver.trans_rdm12(
Lci[iroot], ci[iroot], ncas, nelecas)
casdm1 = (casdm1 * weights[:, None, None]).sum(0)
casdm2 = (casdm2 * weights[:, None, None, None, None]).sum(0)
casdm1 += casdm1.transpose(1, 0)
casdm2 += casdm2.transpose(1, 0, 3, 2)
# gfock = Generalized Fock, Adv. Chem. Phys., 69, 63
dm_core = np.dot(mo_core, mo_core.T) * 2
dm_cas = reduce(np.dot, (mo_cas, casdm1, mo_cas.T))
aapa = ao2mo.kernel(mol, (mo_cas, mo_cas, mo_coeff, mo_cas), compact=False)
aapa = aapa.reshape(ncas, ncas, nmo, ncas)
vj, vk = mc._scf.get_jk(mol, (dm_core, dm_cas))
h1 = mc.get_hcore()
vhf_c = vj[0] - vk[0] * .5
vhf_a = vj[1] - vk[1] * .5
# MRH: delete h1 + vhf_c from the first line below (core and core-core stuff)
# Also extend gfock to span the whole space
gfock = np.zeros_like(dm_cas)
gfock[:, :nocc] = reduce(np.dot, (mo_coeff.T, vhf_a, mo_occ)) * 2
gfock[:, ncore:nocc] = reduce(np.dot,
(mo_coeff.T, h1 + vhf_c, mo_cas, casdm1))
gfock[:, ncore:nocc] += np.einsum('uvpw,vuwt->pt', aapa, casdm2)
dme0 = reduce(np.dot, (mo_coeff, (gfock + gfock.T) * .5, mo_coeff.T))
aapa = vj = vk = vhf_c = vhf_a = h1 = gfock = None
vhf1c, vhf1a = mf_grad.get_veff(mol, (dm_core, dm_cas))
hcore_deriv = mf_grad.hcore_generator(mol)
s1 = mf_grad.get_ovlp(mol)
diag_idx = np.arange(nao)
diag_idx = diag_idx * (diag_idx + 1) // 2 + diag_idx
casdm2_cc = casdm2 + casdm2.transpose(0, 1, 3, 2)
dm2buf = ao2mo._ao2mo.nr_e2(casdm2_cc.reshape(ncas**2, ncas**2), mo_cas.T,
(0, nao, 0, nao)).reshape(ncas**2, nao, nao)
dm2buf = lib.pack_tril(dm2buf)
dm2buf[:, diag_idx] *= .5
dm2buf = dm2buf.reshape(ncas, ncas, nao_pair)
casdm2 = casdm2_cc = None
if atmlst is None:
atmlst = range(mol.natm)
aoslices = mol.aoslice_by_atom()
de_hcore = np.zeros((len(atmlst), 3))
de_renorm = np.zeros((len(atmlst), 3))
de_eri = np.zeros((len(atmlst), 3))
de = np.zeros((len(atmlst), 3))
max_memory = mc.max_memory - lib.current_memory()[0]
blksize = int(max_memory * .9e6 / 8 /
((aoslices[:, 3] - aoslices[:, 2]).max() * nao_pair))
blksize = min(nao, max(2, blksize))
for k, ia in enumerate(atmlst):
shl0, shl1, p0, p1 = aoslices[ia]
h1ao = hcore_deriv(ia)
# MRH: dm1 -> dm_cas in the line below
de_hcore[k] += np.einsum('xij,ij->x', h1ao, dm_cas)
de_renorm[k] -= np.einsum('xij,ij->x', s1[:, p0:p1], dme0[p0:p1]) * 2
q1 = 0
for b0, b1, nf in _shell_prange(mol, 0, mol.nbas, blksize):
q0, q1 = q1, q1 + nf
dm2_ao = lib.einsum('ijw,pi,qj->pqw', dm2buf, mo_cas[p0:p1],
mo_cas[q0:q1])
shls_slice = (shl0, shl1, b0, b1, 0, mol.nbas, 0, mol.nbas)
eri1 = mol.intor('int2e_ip1',
comp=3,
aosym='s2kl',
shls_slice=shls_slice).reshape(
3, p1 - p0, nf, nao_pair)
de_eri[k] -= np.einsum('xijw,ijw->x', eri1, dm2_ao) * 2
eri1 = None
# MRH: dm1 -> dm_cas in the line below. Also eliminate core-core terms
de_eri[k] += np.einsum('xij,ij->x', vhf1c[:, p0:p1], dm_cas[p0:p1]) * 2
de_eri[k] += np.einsum('xij,ij->x', vhf1a[:, p0:p1],
dm_core[p0:p1]) * 2
lib.logger.debug(mc, "CI lagrange hcore component:\n{}".format(de_hcore))
lib.logger.debug(mc, "CI lagrange renorm component:\n{}".format(de_renorm))
lib.logger.debug(mc, "CI lagrange eri component:\n{}".format(de_eri))
de = de_hcore + de_renorm + de_eri
return de
def kernel(mp, t2, atmlst=None, mf_grad=None, verbose=logger.INFO):
if mf_grad is None: mf_grad = mp._scf.nuc_grad_method()
log = logger.new_logger(mp, verbose)
time0 = time.clock(), time.time()
log.debug('Build ump2 rdm1 intermediates')
d1 = ump2._gamma1_intermediates(mp, t2)
time1 = log.timer_debug1('rdm1 intermediates', *time0)
log.debug('Build ump2 rdm2 intermediates')
mol = mp.mol
with_frozen = not (mp.frozen is None or mp.frozen is 0)
moidx = mp.get_frozen_mask()
OA_a, VA_a, OF_a, VF_a = mp2_grad._index_frozen_active(moidx[0], mp.mo_occ[0])
OA_b, VA_b, OF_b, VF_b = mp2_grad._index_frozen_active(moidx[1], mp.mo_occ[1])
orboa = mp.mo_coeff[0][:,OA_a]
orbva = mp.mo_coeff[0][:,VA_a]
orbob = mp.mo_coeff[1][:,OA_b]
orbvb = mp.mo_coeff[1][:,VA_b]
nao, nocca = orboa.shape
nvira = orbva.shape[1]
noccb = orbob.shape[1]
nvirb = orbvb.shape[1]
# Partially transform MP2 density matrix and hold it in memory
# The rest transformation are applied during the contraction to ERI integrals
t2aa, t2ab, t2bb = t2
part_dm2aa = _ao2mo.nr_e2(t2aa.reshape(nocca**2,nvira**2),
numpy.asarray(orbva.T, order='F'), (0,nao,0,nao),
's1', 's1').reshape(nocca,nocca,nao,nao)
part_dm2bb = _ao2mo.nr_e2(t2bb.reshape(noccb**2,nvirb**2),
numpy.asarray(orbvb.T, order='F'), (0,nao,0,nao),
's1', 's1').reshape(noccb,noccb,nao,nao)
part_dm2ab = lib.einsum('ijab,pa,qb->ipqj', t2ab, orbva, orbvb)
part_dm2aa = (part_dm2aa.transpose(0,2,3,1) -
part_dm2aa.transpose(0,3,2,1)) * .5
part_dm2bb = (part_dm2bb.transpose(0,2,3,1) -
part_dm2bb.transpose(0,3,2,1)) * .5
hf_dm1a, hf_dm1b = mp._scf.make_rdm1(mp.mo_coeff, mp.mo_occ)
hf_dm1 = hf_dm1a + hf_dm1b
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))
Imata = numpy.zeros((nao,nao))
Imatb = numpy.zeros((nao,nao))
fdm2 = lib.H5TmpFile()
vhf1 = fdm2.create_dataset('vhf1', (len(atmlst),2,3,nao,nao), 'f8')
# 2e AO integrals dot 2pdm
max_memory = max(0, mp.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((2,3,nao,nao))
for b0, b1, nf in mp2_grad._shell_prange(mol, shl0, shl1, blksize):
ip0, ip1 = ip1, ip1 + nf
dm2bufa = lib.einsum('pi,iqrj->pqrj', orboa[ip0:ip1], part_dm2aa)
dm2bufa+= lib.einsum('qi,iprj->pqrj', orboa, part_dm2aa[:,ip0:ip1])
dm2bufa = lib.einsum('pqrj,sj->pqrs', dm2bufa, orboa)
tmp = lib.einsum('pi,iqrj->pqrj', orboa[ip0:ip1], part_dm2ab)
tmp+= lib.einsum('qi,iprj->pqrj', orboa, part_dm2ab[:,ip0:ip1])
dm2bufa+= lib.einsum('pqrj,sj->pqrs', tmp, orbob)
tmp = None
dm2bufa = dm2bufa + dm2bufa.transpose(0,1,3,2)
dm2bufa = lib.pack_tril(dm2bufa.reshape(-1,nao,nao)).reshape(nf,nao,-1)
dm2bufa[:,:,diagidx] *= .5
dm2bufb = lib.einsum('pi,iqrj->pqrj', orbob[ip0:ip1], part_dm2bb)
dm2bufb+= lib.einsum('qi,iprj->pqrj', orbob, part_dm2bb[:,ip0:ip1])
dm2bufb = lib.einsum('pqrj,sj->pqrs', dm2bufb, orbob)
tmp = lib.einsum('iqrj,sj->iqrs', part_dm2ab, orbob[ip0:ip1])
tmp+= lib.einsum('iqrj,sj->iqsr', part_dm2ab[:,:,ip0:ip1], orbob)
dm2bufb+= lib.einsum('pi,iqrs->srpq', orboa, tmp)
tmp = None
dm2bufb = dm2bufb + dm2bufb.transpose(0,1,3,2)
dm2bufb = lib.pack_tril(dm2bufb.reshape(-1,nao,nao)).reshape(nf,nao,-1)
dm2bufb[:,:,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)
Imata += lib.einsum('ipx,iqx->pq', eri0.reshape(nf,nao,-1), dm2bufa)
Imatb += lib.einsum('ipx,iqx->pq', eri0.reshape(nf,nao,-1), dm2bufb)
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, dm2bufa) * 2
de[k] -= numpy.einsum('xijk,ijk->x', eri1, dm2bufb) * 2
dm2bufa = dm2bufb = 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[0,i] -= numpy.einsum('ijkl,il->kj', eri1tmp, hf_dm1a[ip0:ip1])
vhf[1,i] -= numpy.einsum('ijkl,il->kj', eri1tmp, hf_dm1b[ip0:ip1])
vhf[:,i,ip0:ip1] += numpy.einsum('ijkl,kl->ij', eri1tmp, hf_dm1)
vhf[0,i,ip0:ip1] -= numpy.einsum('ijkl,jk->il', eri1tmp, hf_dm1a)
vhf[1,i,ip0:ip1] -= numpy.einsum('ijkl,jk->il', eri1tmp, hf_dm1b)
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)
# Recompute nocc, nvir to include the frozen orbitals and make contraction for
# the 1-particle quantities, see also the kernel function in uccsd_grad module.
mo_a, mo_b = mp.mo_coeff
mo_ea, mo_eb = mp._scf.mo_energy
nao, nmoa = mo_a.shape
nmob = mo_b.shape[1]
nocca = numpy.count_nonzero(mp.mo_occ[0] > 0)
noccb = numpy.count_nonzero(mp.mo_occ[1] > 0)
s0 = mp._scf.get_ovlp()
Imata = reduce(numpy.dot, (mo_a.T, Imata, s0, mo_a)) * -1
Imatb = reduce(numpy.dot, (mo_b.T, Imatb, s0, mo_b)) * -1
dm1a = numpy.zeros((nmoa,nmoa))
dm1b = numpy.zeros((nmob,nmob))
doo, dOO = d1[0]
dvv, dVV = d1[1]
if with_frozen:
dco = Imata[OF_a[:,None],OA_a] / (mo_ea[OF_a,None] - mo_ea[OA_a])
dfv = Imata[VF_a[:,None],VA_a] / (mo_ea[VF_a,None] - mo_ea[VA_a])
dm1a[OA_a[:,None],OA_a] = (doo + doo.T) * .5
dm1a[OF_a[:,None],OA_a] = dco
dm1a[OA_a[:,None],OF_a] = dco.T
dm1a[VA_a[:,None],VA_a] = (dvv + dvv.T) * .5
dm1a[VF_a[:,None],VA_a] = dfv
dm1a[VA_a[:,None],VF_a] = dfv.T
dco = Imatb[OF_b[:,None],OA_b] / (mo_eb[OF_b,None] - mo_eb[OA_b])
dfv = Imatb[VF_b[:,None],VA_b] / (mo_eb[VF_b,None] - mo_eb[VA_b])
dm1b[OA_b[:,None],OA_b] = (dOO + dOO.T) * .5
dm1b[OF_b[:,None],OA_b] = dco
dm1b[OA_b[:,None],OF_b] = dco.T
dm1b[VA_b[:,None],VA_b] = (dVV + dVV.T) * .5
dm1b[VF_b[:,None],VA_b] = dfv
dm1b[VA_b[:,None],VF_b] = dfv.T
else:
dm1a[:nocca,:nocca] = (doo + doo.T) * .5
dm1a[nocca:,nocca:] = (dvv + dvv.T) * .5
dm1b[:noccb,:noccb] = (dOO + dOO.T) * .5
dm1b[noccb:,noccb:] = (dVV + dVV.T) * .5
dm1 = (reduce(numpy.dot, (mo_a, dm1a, mo_a.T)),
reduce(numpy.dot, (mo_b, dm1b, mo_b.T)))
vhf = mp._scf.get_veff(mp.mol, dm1)
Xvo = reduce(numpy.dot, (mo_a[:,nocca:].T, vhf[0], mo_a[:,:nocca]))
XVO = reduce(numpy.dot, (mo_b[:,noccb:].T, vhf[1], mo_b[:,:noccb]))
Xvo+= Imata[:nocca,nocca:].T - Imata[nocca:,:nocca]
XVO+= Imatb[:noccb,noccb:].T - Imatb[noccb:,:noccb]
dm1_resp = _response_dm1(mp, (Xvo,XVO))
dm1a += dm1_resp[0]
dm1b += dm1_resp[1]
time1 = log.timer_debug1('response_rdm1 intermediates', *time1)
Imata[nocca:,:nocca] = Imata[:nocca,nocca:].T
Imatb[noccb:,:noccb] = Imatb[:noccb,noccb:].T
im1 = reduce(numpy.dot, (mo_a, Imata, mo_a.T))
im1+= reduce(numpy.dot, (mo_b, Imatb, mo_b.T))
time1 = log.timer_debug1('response_rdm1', *time1)
log.debug('h1 and JK1')
hcore_deriv = mf_grad.hcore_generator(mol)
s1 = mf_grad.get_ovlp(mol)
zeta = (mo_ea[:,None] + mo_ea) * .5
zeta[nocca:,:nocca] = mo_ea[:nocca]
zeta[:nocca,nocca:] = mo_ea[:nocca].reshape(-1,1)
zeta_a = reduce(numpy.dot, (mo_a, zeta*dm1a, mo_a.T))
zeta = (mo_eb[:,None] + mo_eb) * .5
zeta[noccb:,:noccb] = mo_eb[:noccb]
zeta[:noccb,noccb:] = mo_eb[:noccb].reshape(-1,1)
zeta_b = reduce(numpy.dot, (mo_b, zeta*dm1b, mo_b.T))
dm1 = (reduce(numpy.dot, (mo_a, dm1a, mo_a.T)),
reduce(numpy.dot, (mo_b, dm1b, mo_b.T)))
vhf_s1occ = mp._scf.get_veff(mol, (dm1[0]+dm1[0].T, dm1[1]+dm1[1].T))
p1a = numpy.dot(mo_a[:,:nocca], mo_a[:,:nocca].T)
p1b = numpy.dot(mo_b[:,:noccb], mo_b[:,:noccb].T)
vhf_s1occ = (reduce(numpy.dot, (p1a, vhf_s1occ[0], p1a)) +
reduce(numpy.dot, (p1b, vhf_s1occ[1], p1b))) * .5
time1 = log.timer_debug1('h1 and JK1', *time1)
# Hartree-Fock part contribution
dm1pa = hf_dm1a + dm1[0]*2
dm1pb = hf_dm1b + dm1[1]*2
dm1 = dm1[0] + dm1[1] + hf_dm1
zeta_a += rhf_grad.make_rdm1e(mo_ea, mo_a, mp.mo_occ[0])
zeta_b += rhf_grad.make_rdm1e(mo_eb, mo_b, mp.mo_occ[1])
zeta = zeta_a + zeta_b
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,0], dm1pa)
de[k] -= numpy.einsum('xij,ij->x', vhf1[k,1], dm1pb)
de += mf_grad.grad_nuc(mol)
log.timer('%s gradients' % mp.__class__.__name__, *time0)
return de
def kernel(mycc, t1=None, t2=None, l1=None, l2=None, eris=None, atmlst=None,
mf_grad=None, d1=None, d2=None, verbose=logger.INFO):
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
if mf_grad is None: mf_grad = mycc._scf.nuc_grad_method()
log = logger.new_logger(mycc, verbose)
time0 = time.clock(), time.time()
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 = mycc.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 mycc.frozen is 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')
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 += mf_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)
de += mf_grad.grad_nuc(mol, atmlst)
log.timer('%s gradients' % mycc.__class__.__name__, *time0)
return de
def Lorb_dot_dgorb_dx(Lorb,
mc,
mo_coeff=None,
ci=None,
atmlst=None,
mf_grad=None,
eris=None,
verbose=None):
''' Modification of single-state CASSCF electronic energy nuclear gradient to compute instead
the orbital Lagrange term nuclear gradient:
sum_pq Lorb_pq d2_Ecas/d_lambda d_kpq
This involves the effective density matrices
~D_pq = L_pr*D_rq + L_qr*D_pr
~d_pqrs = L_pt*d_tqrs + L_rt*d_pqts + L_qt*d_ptrs + L_st*d_pqrt
(NB: L_pq = -L_qp)
'''
# dmo = smoT.dao.smo
# dao = mo.dmo.moT
t0 = (logger.process_clock(), logger.perf_counter())
if mo_coeff is None: mo_coeff = mc.mo_coeff
if ci is None: ci = mc.ci
if mf_grad is None: mf_grad = mc._scf.nuc_grad_method()
if mc.frozen is not None:
raise NotImplementedError
mol = mc.mol
ncore = mc.ncore
ncas = mc.ncas
nocc = ncore + ncas
nelecas = mc.nelecas
nao, nmo = mo_coeff.shape
nao_pair = nao * (nao + 1) // 2
mo_core = mo_coeff[:, :ncore]
mo_cas = mo_coeff[:, ncore:nocc]
# MRH: new 'effective' MO coefficients including contraction from the Lagrange multipliers
moL_coeff = np.dot(mo_coeff, Lorb)
s0_inv = np.dot(mo_coeff, mo_coeff.T)
moL_core = moL_coeff[:, :ncore]
moL_cas = moL_coeff[:, ncore:nocc]
# MRH: these SHOULD be state-averaged! Use the actual sacasscf object!
casdm1, casdm2 = mc.fcisolver.make_rdm12(ci, ncas, nelecas)
# gfock = Generalized Fock, Adv. Chem. Phys., 69, 63
dm_core = np.dot(mo_core, mo_core.T) * 2
dm_cas = reduce(np.dot, (mo_cas, casdm1, mo_cas.T))
# MRH: new density matrix terms
dmL_core = np.dot(moL_core, mo_core.T) * 2
dmL_cas = reduce(np.dot, (moL_cas, casdm1, mo_cas.T))
dmL_core += dmL_core.T
dmL_cas += dmL_cas.T
dm1 = dm_core + dm_cas
dm1L = dmL_core + dmL_cas
# MRH: wrap the integral instead of the density matrix.
# g_prst*~d_qrst = (g_pust*L_ur + g_prut*L_us + g_prsu*L_ut)*d_qrst + g_prst*L_uq*d_urst
# = 'aapaL'_prst*d_qrst [ERI TERM 1]
# = 'aapa'_prst*L_uq*d_urst [ERI TERM 2]
aapa = np.zeros((ncas, ncas, nmo, ncas), dtype=dm_cas.dtype)
aapaL = np.zeros((ncas, ncas, nmo, ncas), dtype=dm_cas.dtype)
for i in range(nmo):
jbuf = eris.ppaa[i]
kbuf = eris.papa[i]
aapa[:, :, i, :] = jbuf[ncore:nocc, :, :].transpose(1, 2, 0)
aapaL[:, :, i, :] += np.tensordot(jbuf,
Lorb[:, ncore:nocc],
axes=((0), (0)))
kbuf = np.tensordot(kbuf, Lorb[:, ncore:nocc],
axes=((1), (0))).transpose(1, 2, 0)
aapaL[:, :, i, :] += kbuf + kbuf.transpose(1, 0, 2)
# MRH: new vhf terms
vj, vk = mc._scf.get_jk(mol, (dm_core, dm_cas))
vjL, vkL = mc._scf.get_jk(mol, (dmL_core, dmL_cas))
h1 = mc.get_hcore()
vhf_c = vj[0] - vk[0] * .5
vhf_a = vj[1] - vk[1] * .5
vhfL_c = vjL[0] - vkL[0] * .5
vhfL_a = vjL[1] - vkL[1] * .5
gfock = np.dot(h1, dm1L) # h1e
gfock += np.dot((vhf_c + vhf_a),
dmL_core) # core-core and active-core, 2nd 1RDM linked
gfock += np.dot((vhfL_c + vhfL_a),
dm_core) # core-core and active-core, 1st 1RDM linked
gfock += np.dot(vhfL_c, dm_cas) # core-active, 1st 1RDM linked
gfock += np.dot(vhf_c, dmL_cas) # core-active, 2nd 1RDM linked
gfock = np.dot(
s0_inv,
gfock) # Definition in MO's; going (AO->MO->AO) incurs inverse ovlp
# [ERI TERM 1]
gfock += reduce(
np.dot,
(mo_coeff, np.einsum('uviw,uvtw->it', aapaL, casdm2), mo_cas.T))
# [ERI TERM 2]
gfock += reduce(
np.dot,
(mo_coeff, np.einsum('uviw,vuwt->it', aapa, casdm2), moL_cas.T))
dme0 = (gfock +
gfock.T) / 2 # This transpose is for the overlap matrix later on
aapa = vj = vk = vhf_c = vhf_a = None
vj, vk = mf_grad.get_jk(mol, (dm_core, dm_cas, dmL_core, dmL_cas))
vhf1c, vhf1a, vhf1cL, vhf1aL = vj - vk * 0.5
hcore_deriv = mf_grad.hcore_generator(mol)
s1 = mf_grad.get_ovlp(mol)
diag_idx = np.arange(nao)
diag_idx = diag_idx * (diag_idx + 1) // 2 + diag_idx
casdm2_cc = casdm2 + casdm2.transpose(0, 1, 3, 2)
dm2buf = ao2mo._ao2mo.nr_e2(casdm2_cc.reshape(ncas**2, ncas**2), mo_cas.T,
(0, nao, 0, nao)).reshape(ncas**2, nao, nao)
# MRH: contract the final two indices of the active-active 2RDM with L as you change to AOs
# note tensordot always puts indices in the order of the arguments.
dm2Lbuf = np.zeros((ncas**2, nmo, nmo))
# MRH: The second line below transposes the L; the third line transposes the derivative
# Both the L and the derivative have to explore all indices
Lcasdm2 = np.tensordot(Lorb[:, ncore:nocc], casdm2,
axes=(1, 2)).transpose(1, 2, 0, 3)
dm2Lbuf[:, :, ncore:nocc] = Lcasdm2.reshape(ncas**2, nmo, ncas)
Lcasdm2 = np.tensordot(Lorb[:, ncore:nocc], casdm2,
axes=(1, 3)).transpose(1, 2, 3, 0)
dm2Lbuf[:, ncore:nocc, :] += Lcasdm2.reshape(ncas**2, ncas, nmo)
Lcasdm2 = None
dm2Lbuf += dm2Lbuf.transpose(0, 2, 1)
dm2Lbuf = np.ascontiguousarray(dm2Lbuf)
dm2Lbuf = ao2mo._ao2mo.nr_e2(dm2Lbuf.reshape(ncas**2, nmo**2), mo_coeff.T,
(0, nao, 0, nao)).reshape(ncas**2, nao, nao)
dm2buf = lib.pack_tril(dm2buf)
dm2buf[:, diag_idx] *= .5
dm2buf = dm2buf.reshape(ncas, ncas, nao_pair)
dm2Lbuf = lib.pack_tril(dm2Lbuf)
dm2Lbuf[:, diag_idx] *= .5
dm2Lbuf = dm2Lbuf.reshape(ncas, ncas, nao_pair)
if atmlst is None:
atmlst = list(range(mol.natm))
aoslices = mol.aoslice_by_atom()
de_hcore = np.zeros((len(atmlst), 3))
de_renorm = np.zeros((len(atmlst), 3))
de_eri = np.zeros((len(atmlst), 3))
de = np.zeros((len(atmlst), 3))
max_memory = mc.max_memory - lib.current_memory()[0]
blksize = int(max_memory * .9e6 / 8 /
(4 * (aoslices[:, 3] - aoslices[:, 2]).max() * nao_pair))
# MRH: 3 components of eri array and 1 density matrix array:
# FOUR arrays of this size are required!
blksize = min(nao, max(2, blksize))
logger.info(
mc,
'SA-CASSCF Lorb_dot_dgorb memory remaining for eri manipulation: %f MB; using'
' blocksize = %d', max_memory, blksize)
t0 = logger.timer(mc, 'SA-CASSCF Lorb_dot_dgorb 1-electron part', *t0)
for k, ia in enumerate(atmlst):
shl0, shl1, p0, p1 = aoslices[ia]
h1ao = hcore_deriv(ia)
# MRH: h1e and Feff terms
de_hcore[k] += np.einsum('xij,ij->x', h1ao, dm1L)
de_renorm[k] -= np.einsum('xij,ij->x', s1[:, p0:p1], dme0[p0:p1]) * 2
q1 = 0
for b0, b1, nf in _shell_prange(mol, 0, mol.nbas, blksize):
q0, q1 = q1, q1 + nf
dm2_ao = lib.einsum('ijw,pi,qj->pqw', dm2Lbuf, mo_cas[p0:p1],
mo_cas[q0:q1])
# MRH: contract first two indices of active-active 2RDM with L as you go MOs -> AOs
dm2_ao += lib.einsum('ijw,pi,qj->pqw', dm2buf, moL_cas[p0:p1],
mo_cas[q0:q1])
dm2_ao += lib.einsum('ijw,pi,qj->pqw', dm2buf, mo_cas[p0:p1],
moL_cas[q0:q1])
shls_slice = (shl0, shl1, b0, b1, 0, mol.nbas, 0, mol.nbas)
eri1 = mol.intor('int2e_ip1',
comp=3,
aosym='s2kl',
shls_slice=shls_slice).reshape(
3, p1 - p0, nf, nao_pair)
# MRH: I still don't understand why there is a minus here!
de_eri[k] -= np.einsum('xijw,ijw->x', eri1, dm2_ao) * 2
eri1 = dm2_ao = None
t0 = logger.timer(
mc, 'SA-CASSCF Lorb_dot_dgorb atom {} ({},{}|{})'.format(
ia, p1 - p0, nf, nao_pair), *t0)
# MRH: core-core and core-active 2RDM terms
de_eri[k] += np.einsum('xij,ij->x', vhf1c[:, p0:p1], dm1L[p0:p1]) * 2
de_eri[k] += np.einsum('xij,ij->x', vhf1cL[:, p0:p1], dm1[p0:p1]) * 2
# MRH: active-core 2RDM terms
de_eri[k] += np.einsum('xij,ij->x', vhf1a[:, p0:p1],
dmL_core[p0:p1]) * 2
de_eri[k] += np.einsum('xij,ij->x', vhf1aL[:, p0:p1],
dm_core[p0:p1]) * 2
# MRH: deleted the nuclear-nuclear part to avoid double-counting
# lesson learned from debugging - mol.intor computes -1 * the derivative and only
# for one index
# on the other hand, mf_grad.hcore_generator computes the actual derivative of
# h1 for both indices and with the correct sign
logger.debug(mc, "Orb lagrange hcore component:\n{}".format(de_hcore))
logger.debug(mc, "Orb lagrange renorm component:\n{}".format(de_renorm))
logger.debug(mc, "Orb lagrange eri component:\n{}".format(de_eri))
de = de_hcore + de_renorm + de_eri
return de
def make_rdm1_with_orbital_response(mp):
import time
from pyscf import lib
from pyscf.grad.mp2 import _response_dm1, _index_frozen_active, _shell_prange
from pyscf.mp import mp2
from pyscf.ao2mo import _ao2mo
log = lib.logger.new_logger(mp)
time0 = time.clock(), time.time()
mol = mp.mol
log.debug('Build mp2 rdm1 intermediates')
d1 = mp2._gamma1_intermediates(mp, mp.t2)
doo, dvv = d1
time1 = log.timer_debug1('rdm1 intermediates', *time0)
with_frozen = not (mp.frozen is None or mp.frozen is 0)
OA, VA, OF, VF = _index_frozen_active(mp.get_frozen_mask(), mp.mo_occ)
orbo = mp.mo_coeff[:, OA]
orbv = mp.mo_coeff[:, VA]
nao, nocc = orbo.shape
nvir = orbv.shape[1]
# Partially transform MP2 density matrix and hold it in memory
# The rest transformation are applied during the contraction to ERI integrals
part_dm2 = _ao2mo.nr_e2(mp.t2.reshape(nocc**2, nvir**2),
numpy.asarray(orbv.T, order='F'), (0, nao, 0, nao),
's1', 's1').reshape(nocc, nocc, nao, nao)
part_dm2 = (part_dm2.transpose(0, 2, 3, 1) * 4 -
part_dm2.transpose(0, 3, 2, 1) * 2)
offsetdic = mol.offset_nr_by_atom()
diagidx = numpy.arange(nao)
diagidx = diagidx * (diagidx + 1) // 2 + diagidx
Imat = numpy.zeros((nao, nao))
# 2e AO integrals dot 2pdm
max_memory = max(0, mp.max_memory - lib.current_memory()[0])
blksize = max(1, int(max_memory * .9e6 / 8 / (nao**3 * 2.5)))
for ia in range(mol.natm):
shl0, shl1, p0, p1 = offsetdic[ia]
ip1 = p0
for b0, b1, nf in _shell_prange(mol, shl0, shl1, blksize):
ip0, ip1 = ip1, ip1 + nf
dm2buf = lib.einsum('pi,iqrj->pqrj', orbo[ip0:ip1], part_dm2)
dm2buf += lib.einsum('qi,iprj->pqrj', orbo, part_dm2[:, ip0:ip1])
dm2buf = lib.einsum('pqrj,sj->pqrs', dm2buf, orbo)
dm2buf = dm2buf + dm2buf.transpose(0, 1, 3, 2)
dm2buf = lib.pack_tril(dm2buf.reshape(-1, nao,
nao)).reshape(nf, nao, -1)
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
dm2buf = None
time1 = log.timer_debug1('2e-part grad of atom %d' % ia, *time1)
# Recompute nocc, nvir to include the frozen orbitals and make contraction for
# the 1-particle quantities, see also the kernel function in ccsd_grad module.
mo_coeff = mp.mo_coeff
mo_energy = mp._scf.mo_energy
nao, nmo = mo_coeff.shape
nocc = numpy.count_nonzero(mp.mo_occ > 0)
Imat = reduce(numpy.dot,
(mo_coeff.T, Imat, mp._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 = mp._scf.get_veff(mp.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(mp, Xvo)
# Transform to AO basis
dm1 = reduce(numpy.dot, (mo_coeff, dm1mo, mo_coeff.T))
dm1 += mp._scf.make_rdm1(mp.mo_coeff, mp.mo_occ)
return dm1
def kernel(mc, mo_coeff=None, ci=None, atmlst=None, mf_grad=None,
verbose=None):
if mo_coeff is None: mo_coeff = mc.mo_coeff
if ci is None: ci = mc.ci
if mf_grad is None: mf_grad = mc._scf.nuc_grad_method()
if mc.frozen is not None:
raise NotImplementedError
mol = mc.mol
ncore = mc.ncore
ncas = mc.ncas
nocc = ncore + ncas
nelecas = mc.nelecas
nao, nmo = mo_coeff.shape
nao_pair = nao * (nao+1) // 2
mo_occ = mo_coeff[:,:nocc]
mo_core = mo_coeff[:,:ncore]
mo_cas = mo_coeff[:,ncore:nocc]
casdm1, casdm2 = mc.fcisolver.make_rdm12(ci, ncas, nelecas)
# gfock = Generalized Fock, Adv. Chem. Phys., 69, 63
dm_core = numpy.dot(mo_core, mo_core.T) * 2
dm_cas = reduce(numpy.dot, (mo_cas, casdm1, mo_cas.T))
aapa = ao2mo.kernel(mol, (mo_cas, mo_cas, mo_occ, mo_cas), compact=False)
aapa = aapa.reshape(ncas,ncas,nocc,ncas)
vj, vk = mc._scf.get_jk(mol, (dm_core, dm_cas))
h1 = mc.get_hcore()
vhf_c = vj[0] - vk[0] * .5
vhf_a = vj[1] - vk[1] * .5
gfock = reduce(numpy.dot, (mo_occ.T, h1 + vhf_c + vhf_a, mo_occ)) * 2
gfock[:,ncore:nocc] = reduce(numpy.dot, (mo_occ.T, h1 + vhf_c, mo_cas, casdm1))
gfock[:,ncore:nocc] += numpy.einsum('uviw,vuwt->it', aapa, casdm2)
dme0 = reduce(numpy.dot, (mo_occ, (gfock+gfock.T)*.5, mo_occ.T))
aapa = vj = vk = vhf_c = vhf_a = h1 = gfock = None
dm1 = dm_core + dm_cas
vhf1c, vhf1a = mf_grad.get_veff(mol, (dm_core, dm_cas))
hcore_deriv = mf_grad.hcore_generator(mol)
s1 = mf_grad.get_ovlp(mol)
diag_idx = numpy.arange(nao)
diag_idx = diag_idx * (diag_idx+1) // 2 + diag_idx
casdm2_cc = casdm2 + casdm2.transpose(0,1,3,2)
dm2buf = ao2mo._ao2mo.nr_e2(casdm2_cc.reshape(ncas**2,ncas**2), mo_cas.T,
(0, nao, 0, nao)).reshape(ncas**2,nao,nao)
dm2buf = lib.pack_tril(dm2buf)
dm2buf[:,diag_idx] *= .5
dm2buf = dm2buf.reshape(ncas,ncas,nao_pair)
casdm2 = casdm2_cc = None
if atmlst is None:
atmlst = range(mol.natm)
aoslices = mol.aoslice_by_atom()
de = numpy.zeros((len(atmlst),3))
max_memory = mc.max_memory - lib.current_memory()[0]
blksize = int(max_memory*.9e6/8 / ((aoslices[:,3]-aoslices[:,2]).max()*nao_pair))
blksize = min(nao, max(2, blksize))
for k, ia in enumerate(atmlst):
shl0, shl1, p0, p1 = aoslices[ia]
h1ao = hcore_deriv(ia)
de[k] += numpy.einsum('xij,ij->x', h1ao, dm1)
de[k] -= numpy.einsum('xij,ij->x', s1[:,p0:p1], dme0[p0:p1]) * 2
q1 = 0
for b0, b1, nf in _shell_prange(mol, 0, mol.nbas, blksize):
q0, q1 = q1, q1 + nf
dm2_ao = lib.einsum('ijw,pi,qj->pqw', dm2buf, mo_cas[p0:p1], mo_cas[q0:q1])
shls_slice = (shl0,shl1,b0,b1,0,mol.nbas,0,mol.nbas)
eri1 = mol.intor('int2e_ip1', comp=3, aosym='s2kl',
shls_slice=shls_slice).reshape(3,p1-p0,nf,nao_pair)
de[k] -= numpy.einsum('xijw,ijw->x', eri1, dm2_ao) * 2
eri1 = None
de[k] += numpy.einsum('xij,ij->x', vhf1c[:,p0:p1], dm1[p0:p1]) * 2
de[k] += numpy.einsum('xij,ij->x', vhf1a[:,p0:p1], dm_core[p0:p1]) * 2
de += mf_grad.grad_nuc(mol, atmlst)
return de
