Method coverage#

pyscfad re-implements a subset of pyscf’s methods so that they can be differentiated. This page lists what that subset currently is, so that a missing method can be told apart from a bug.

The tables below are a snapshot of pyscfad 0.3.4. Anything not listed either has no differentiable counterpart in pyscfad, or is only usable through mol.to_pyscf() / mf.to_pyscf(), which strips the traced attributes and returns a plain pyscf object.

Most of the methods below are built on the legacy gto.Mole. That path is gradually being transitioned to the fully jittable gto.MoleLite path (the *Lite classes listed in the tables), and new methods may not be added to the legacy path. New code is encouraged to use gto.MoleLite and its *Lite methods wherever a counterpart exists, falling back to the legacy gto.Mole only for the methods that have not been ported yet.

Mean field#

Method	Class	Notes
RHF	`scf.RHF`
UHF	`scf.UHF`
ROHF	`scf.ROHF`
GHF	`scf.GHF`
RKS	`dft.RKS`
UKS	`dft.UKS`
Density fitting	`mf.density_fit()`, `df.DF`	RHF, UHF, ROHF, GHF, RKS, UKS
Fully jittable RHF	`scf.hf_lite.SCFLite`	used with `gto.MoleLite`

ROKS is not implemented.

Correlated methods#

Method	Class	Notes
RMP2	`mp.RMP2`	`mp.MP2(mf)` dispatches on the reference
UMP2	`mp.UMP2`
DF-RMP2	`mp.dfmp2.MP2`
RCCSD	`cc.RCCSD`
DF-RCCSD	`cc.dfccsd.RCCSD`
RCCSD(T)	`mycc.ccsd_t()`	`cc.ccsd_t` (C kernels) or `cc.ccsd_t_slow`
RDCSD	`cc.dfdcsd.RDCSD`
Fully jittable RCCSD	`cc.RCCSDLite`	used with `scf.hf_lite.SCFLite`
LNO-MP2 / LNO-CCSD / LNO-CCSD(T)	`lno.LNOMP2`, `lno.LNOCCSD`, `lno.LNOCCSD_T`
direct RPA	`gw.rpa.RPA`
CIS	`tdscf.CIS`	RHF reference
FCI	`fci.fci_slow`	functions only, no class

GMP2, and the unrestricted and generalized coupled-cluster methods (UCCSD, UCCSD(T), GCCSD, GCCSD(T)) are not implemented. Neither are CISD, CASSCF/CASCI, and the GW methods other than direct RPA.

Other modules#

Module	Contents
`pyscfad.lo`	Boys (`lo.boys.Boys`) and Pipek-Mezey (`lo.pipek.PM`) localization
`pyscfad.prop`	RHF polarizability (`prop.polarizability.rhf`), thermochemistry
`pyscfad.geomopt`	geometry optimization through `geometric`
`pyscfad.xtb`	GFN1-xTB, molecular and periodic, with QM/MM
`pyscfad.ml`	padded `ml.gto.MolePad` / `ml.scf.SCFPad` for batched machine learning

Periodic boundary conditions#

pyscfad.pbc mirrors a smaller part of pyscf.pbc:

Method	Class	Notes
RHF (gamma point)	`pbc.scf.RHF`
KRHF	`pbc.scf.KRHF`
RKS (gamma point)	`pbc.dft.RKS`
KRKS	`pbc.dft.KRKS`
FFTDF	`pbc.df.FFTDF`	the only density fitting scheme
Fully jittable KRHF	`pbc.scf.khf_lite.KSCFLite`	used with `pbc.gto.CellLite`

Unrestricted periodic mean fields, GDF/MDF/RSDF, and the periodic correlated methods are not implemented.

Derivatives#

Where a method is listed above, first derivatives with respect to the traced attributes of gto.Mole (coords, exp, ctr_coeff) are supported. Higher order derivatives are covered by the test suite for the molecular integrals, the SCF and DFT methods, and RCCSD; elsewhere they may work but are not tested.

Warning

Basis parameter derivatives on the legacy gto.Mole path are not taken with respect to the raw parameters of the basis set. exp and ctr_coeff are read out of the already built mol._env, into which pyscf has folded the primitive and contracted-AO normalization: ctr_coeff holds the normalized contraction coefficients rather than the ones in the basis set definition, and the exp derivative is taken with those coefficients held fixed, so it misses the dependence of the normalization factors on the exponent. gto.MoleLite instead builds the basis inside the traced computation, and so differentiates the raw exponents and coefficients.