Source code for vibeqc.periodic_uhf_ewald

"""Phase 15a: Γ-point periodic UHF SCF driver using the composed
Ewald-3D Coulomb dispatch.

Open-shell counterpart of :mod:`vibeqc.periodic_rhf_ewald`. Carries
two density matrices ``D_a``, ``D_b`` (one-particle, no factor 2)
with the standard UHF Fock construction

    F_a = H_core + J(D_a + D_b, w)  -  K(D_a)
    F_b = H_core + J(D_a + D_b, w)  -  K(D_b)

where the Hartree J uses the w-invariant Ewald split from
:func:`vibeqc.build_j_ewald_3d` (Phase 12e-c-4a) and the per-spin
exchange K comes from the full-range real-space builder
(:func:`vibeqc.build_jk_gamma_molecular_limit` at w = 0). DIIS runs
on the *block-diagonal* error vector ``[e_a, e_b]`` with one rolling
F history per spin -- same convention as the molecular UHF C++
driver.

Occupations from the molecule's ``multiplicity``:
    n_a = (n_e + mult - 1) / 2
    n_b = (n_e - mult + 1) / 2

Reports ``<S^2> = (n_a - n_b)(n_a - n_b + 2)/4
                 + n_b - S_{ij in occ} |<phi_i^a | phi_j^b>|^2``
as the standard spin-contamination diagnostic (UHF tests rely on
matching this to the molecular UHF driver in the molecular limit).

Scope (this commit)
-------------------

  - Γ-only (single k-point at the origin). Multi-k UHF lands in
    Phase 15b on top of the multi-k Ewald RHF substrate.
  - Closed-shell special case (multiplicity = 1) reproduces the RHF
    SCF result to ~µHa -- sanity check against
    :func:`run_rhf_periodic_gamma_ewald3d`.
  - DIIS supported (default on); per-spin damping in the no-DIIS
    fallback.
"""

from __future__ import annotations

import math
from dataclasses import dataclass, field
from typing import List, Optional, Sequence, Tuple, Union

import numpy as np

from ._vibeqc_core import (
    apply_level_shift,
    BasisSet,
    bloch_sum,
    build_jk_gamma_molecular_limit,
    compute_kinetic_lattice,
    compute_nuclear_lattice,
    compute_overlap_lattice,
    CoulombMethod,
    InitialGuess,
    LatticeSumOptions,
    level_shift_at_iter,
    LevelShiftDensity,
    nuclear_repulsion_per_cell,
    PeriodicRHFOptions,
    PeriodicSystem,
    SCFIteration,
    SpinlockMode,
)
from .ewald_composed import make_ewald_3d_gamma_j_builder
from .ewald_j import auto_grid
from .guess import initial_densities_open_shell
from .madelung import (
    madelung_energy_correction_for_lat as _madelung_energy_correction_for_lat,
)
from .mom import reorder_occupied_by_max_overlap as _mom_reorder
from .periodic_rhf_ewald import _canonical_orthogonalizer, _refuse_if_dense_ionic
from .periodic_scf_accelerators import DynamicDamping, PeriodicSCFAccelerator
from .progress import ProgressLogger, resolve_progress
from .scf_divergence import check_scf_divergence

__all__ = [
    "PeriodicUHFEwaldResult",
    "run_uhf_periodic_gamma_ewald3d",
]


[docs] @dataclass class PeriodicUHFEwaldResult: """Result of :func:`run_uhf_periodic_gamma_ewald3d`. Mirrors the molecular ``UHFResult`` shape with a separate a/b block, plus Ewald bookkeeping (``omega``, ``grid_shape``). All matrices are at Γ (no Bloch sums). """ energy: float e_electronic: float e_nuclear: float n_iter: int converged: bool s_squared: float s_squared_ideal: float # a spin mo_energies_alpha: np.ndarray mo_coeffs_alpha: np.ndarray density_alpha: np.ndarray fock_alpha: np.ndarray # b spin mo_energies_beta: np.ndarray mo_coeffs_beta: np.ndarray density_beta: np.ndarray fock_beta: np.ndarray overlap: np.ndarray scf_trace: List[SCFIteration] = field(default_factory=list) omega: float = 0.0 grid_shape: Tuple[int, int, int] = (0, 0, 0)
def _spin_squared( n_alpha: int, n_beta: int, C_alpha: np.ndarray, C_beta: np.ndarray, S: np.ndarray, ) -> float: """<S^2> for a UHF wavefunction. ``S^2_UHF = (n_a - n_b)(n_a - n_b + 2)/4 + n_b - S_{i in occ_a, j in occ_b} |<phi_i^a | phi_j^b>|^2`` Matches the molecular C++ driver's convention. The exact eigenvalue for a pure spin state with multiplicity ``mult`` is ``S(S+1)`` where ``S = (mult - 1)/2``. """ diff = n_alpha - n_beta s2 = 0.25 * diff * (diff + 2) + n_beta if n_alpha == 0 or n_beta == 0: return float(s2) Ca_occ = C_alpha[:, :n_alpha] Cb_occ = C_beta[:, :n_beta] overlap_ab = Ca_occ.T @ S @ Cb_occ s2 -= float(np.sum(np.abs(overlap_ab) ** 2)) return float(s2)
[docs] def run_uhf_periodic_gamma_ewald3d( system: PeriodicSystem, basis: BasisSet, options: Optional[PeriodicRHFOptions] = None, *, omega: float = 0.0, grid_shape: Optional[Union[Tuple[int, int, int], int]] = None, origin: Optional[Sequence[float]] = None, spacing_bohr: float = 0.3, linear_dep_threshold: float = 1e-7, canonical_orth_normalize_diag_first: bool = True, auto_optimize_truncation: bool = True, allow_dense_ionic: bool = False, progress: Union[bool, ProgressLogger, None] = None, verbose: Optional[int] = None, ) -> PeriodicUHFEwaldResult: """Γ-point open-shell UHF SCF with the EWALD_3D Coulomb dispatch. Same interface as :func:`run_rhf_periodic_gamma_ewald3d` but with a/b occupations driven by the molecule's ``multiplicity``. Reuses the RHF driver's ``_canonical_orthogonalizer`` and the shared :class:`PeriodicSCFAccelerator` from ``periodic_scf_accelerators``; the Ewald J / real-space K builds are identical to the closed-shell path. Parameters ---------- system, basis, options See :func:`run_rhf_periodic_gamma_ewald3d`. ``options`` is a :class:`PeriodicRHFOptions` (UHF-specific options haven't diverged from RHF in vibe-qc's Python drivers -- same DIIS, damping, level shift, etc.). omega, grid_shape, origin, spacing_bohr, linear_dep_threshold Ewald + numeric controls. See the RHF driver. """ opts = options if options is not None else PeriodicRHFOptions() # SPINLOCK SPIN_SCHEDULE: two-phase (lock spin for spinlock_iterations # cycles, then release to the target multiplicity from that density). # Delegated before any SCF setup; the two sub-runs re-enter this driver # with spinlock_mode = OFF, so there is no further recursion. if ( getattr(opts, "spinlock_mode", SpinlockMode.OFF) == SpinlockMode.SPIN_SCHEDULE and int(getattr(opts, "spinlock_iterations", 0)) > 0 ): from .spinlock_periodic import run_spin_schedule return run_spin_schedule( lambda sysx, o: run_uhf_periodic_gamma_ewald3d( sysx, basis, o, omega=omega, grid_shape=grid_shape, origin=origin, spacing_bohr=spacing_bohr, linear_dep_threshold=linear_dep_threshold, canonical_orth_normalize_diag_first=canonical_orth_normalize_diag_first, auto_optimize_truncation=auto_optimize_truncation, allow_dense_ionic=allow_dense_ionic, progress=progress, verbose=verbose, ), system, opts, ) lat_opts: LatticeSumOptions = opts.lattice_opts plog = resolve_progress(progress, verbose=verbose) # ---- Force EWALD_3D gauge (gauge consistency; handover F4 2026-06-01) ---- # This driver hard-codes the Hartree J to the Ewald-3D builder, so V_ne # (compute_nuclear_lattice_dispatch) and e_nuc (nuclear_repulsion_per_cell) # MUST share that gauge. Without the force, a default options object # (coulomb_method=DIRECT_TRUNCATED) makes nuclear_repulsion_per_cell return # the molecular 1/d sum and madelung_energy_correction_for_lat the bare-gauge # +a_M.Q_e^2/2L term; those only partially cancel (~0.74 mHa on H2/30-bohr), # so the SCF converged to a non-physical energy with no warning (CLAUDE.md # Sec.7). The closed-shell Γ / multi-k RHF/RKS siblings already force this # (audit F1); extending it here aligns e_nuclear with # run_rhf_periodic_gamma_ewald3d and zeroes the now-redundant Madelung term # (madelung_energy_correction_for_lat returns 0.0 for EWALD_3D). # EWALD_3D V_ne (compute_nuclear_lattice_dispatch) is implemented only for # dim == 3 -- the 1D/2D Ewald variants raise (periodic_v_ne.py). So gate the # force on dim == 3; low-dim cells keep their DIRECT_TRUNCATED gauge (the # historical behaviour for these drivers on 1D/2D chains). if system.dim == 3 and lat_opts.coulomb_method != CoulombMethod.EWALD_3D: plog.info( "coulomb_method forced to EWALD_3D for gauge consistency " f"(was {lat_opts.coulomb_method!r}); this driver's Hartree J " "is Ewald-3D and V_ne / e_nuc must match" ) lat_opts.coulomb_method = CoulombMethod.EWALD_3D # Fail closed if periodic images of the basis overlap enough to make the # molecular-limit energy wrong (CLAUDE.md Sec.7); GDF/BIPOLE/GPW are correct # there. ``allow_dense_ionic=True`` bypasses for mechanics testing only. _refuse_if_dense_ionic(system, basis, lat_opts, allow_dense_ionic) # w must match the nuclear Ewald a (auto-selected from # nuclear_cutoff_bohr in the C++ ewald engine) so the jellium # background terms cancel exactly. Mirrors the override block in # run_rhf_periodic_gamma_ewald3d (commit 49f8ae91). The driver # kwarg ``omega`` is retained for signature parity but overridden # here; users override via ``opts.ewald_omega`` when needed. _ewald_tol = getattr(opts, "ewald_tolerance", 1e-12) _cutoff = getattr(opts, "ewald_cutoff_bohr", lat_opts.nuclear_cutoff_bohr) if omega <= 0.0: _user_omega = getattr(opts, "ewald_omega", None) if _user_omega is not None and float(_user_omega) > 0.0: omega = float(_user_omega) else: from .bipole_ext_el_pole import crystal_default_ewald_alpha V_cell = float(abs(np.linalg.det(np.asarray(system.lattice, dtype=float)))) omega = crystal_default_ewald_alpha(V_cell) lat = np.asarray(system.lattice, dtype=float) if grid_shape is None: grid_shape_t = auto_grid(lat, spacing_bohr) elif isinstance(grid_shape, int): grid_shape_t = (grid_shape, grid_shape, grid_shape) else: grid_shape_t = tuple(int(x) for x in grid_shape) plog.info( f"UHF Gamma EWALD_3D / omega = {float(omega):.3f}, " f"FFT grid {grid_shape_t[0]}x{grid_shape_t[1]}x{grid_shape_t[2]}" ) plog.info(f"basis: {basis.name} ({basis.nbasis} BFs / {basis.nshells} shells)") # Active-settings dump. from .options_dump import dump_active_settings dump_active_settings( plog, [ ("PeriodicRHFOptions", opts), ("LatticeSumOptions", lat_opts), ( "Driver kwargs", { "omega": float(omega), "grid_shape": grid_shape_t, "origin": origin, "spacing_bohr": float(spacing_bohr), "linear_dep_threshold": float(linear_dep_threshold), "canonical_orth_normalize_diag_first": canonical_orth_normalize_diag_first, "auto_optimize_truncation": auto_optimize_truncation, }, ), ], ) if plog.level >= 5: from .scf_log import format_basis_summary plog.write_raw(format_basis_summary(basis)) # ---- Open-shell occupations ------------------------------------------ n_elec = int(system.n_electrons()) mult = int(system.multiplicity) if mult < 1: raise ValueError( f"run_uhf_periodic_gamma_ewald3d: multiplicity must be >= 1, got {mult}" ) if (n_elec + mult - 1) % 2 != 0 or (n_elec - mult + 1) % 2 != 0: raise ValueError( f"run_uhf_periodic_gamma_ewald3d: (n_electrons={n_elec}, " f"multiplicity={mult}) cannot be split into integer " f"a/b occupations." ) n_alpha = (n_elec + mult - 1) // 2 n_beta = (n_elec - mult + 1) // 2 if n_alpha < 0 or n_beta < 0: raise ValueError( f"run_uhf_periodic_gamma_ewald3d: invalid occupations " f"n_a={n_alpha}, n_b={n_beta} for n_e={n_elec}, " f"mult={mult}" ) plog.info( f"UHF Gamma occupations: n_alpha = {n_alpha}, " f"n_beta = {n_beta} (multiplicity = {mult})" ) # ---- Auto-optimise lattice truncation (default ON) ------------------- if auto_optimize_truncation and lat_opts.coulomb_method == CoulombMethod.EWALD_3D: from .eigs_preflight import ( format_truncation_optimization_report, optimize_truncation, ) opt_rep = optimize_truncation(system, basis, lattice_opts=lat_opts) if ( opt_rep.n_evaluations > 1 or opt_rep.optimized_lattice_opts.cutoff_bohr != lat_opts.cutoff_bohr ): plog.write_raw(format_truncation_optimization_report(opt_rep)) if not opt_rep.converged: plog.warn( "auto_optimize_truncation did not converge; SCF will run " "with the last evaluated settings." ) lat_opts = opt_rep.optimized_lattice_opts # ---- One-electron integrals at Γ ------------------------------------- with plog.stage( "integrals_lattice", detail=f"S/T/V at cutoff {lat_opts.cutoff_bohr:.2f} bohr" ): S_lat = compute_overlap_lattice(basis, system, lat_opts) T_lat = compute_kinetic_lattice(basis, system, lat_opts) from .periodic_v_ne import compute_nuclear_lattice_dispatch V_lat = compute_nuclear_lattice_dispatch(basis, system, lat_opts) k_gamma = np.zeros(3) S = np.real(bloch_sum(S_lat, k_gamma)) T = np.real(bloch_sum(T_lat, k_gamma)) V = np.real(bloch_sum(V_lat, k_gamma)) Hcore = T + V S = 0.5 * (S + S.T) Hcore = 0.5 * (Hcore + Hcore.T) # Linear-dependence preflight on S(Γ); see periodic_rhf_ewald.py # for the rationale. from .linear_dependence import scf_preflight_overlap_check scf_preflight_overlap_check( S, plog=plog, label="S(Γ)", basis=basis, ) # Canonical-orth on shared S (a and b share the basis orthogonaliser). X, n_kept = _canonical_orthogonalizer( S, linear_dep_threshold, normalize_diag_first=canonical_orth_normalize_diag_first, ) if max(n_alpha, n_beta) > n_kept: raise RuntimeError( f"run_uhf_periodic_gamma_ewald3d: canonical orthogonalisation " f"kept {n_kept} directions; need >= {max(n_alpha, n_beta)} " f"(n_a={n_alpha}, n_b={n_beta})." ) use_davidson = getattr(opts, "use_davidson", False) dav_opts = getattr(opts, "davidson", None) dav_dim = getattr(opts, "davidson_min_dim", 100) use_dav = use_davidson and S.shape[0] >= dav_dim if use_dav and dav_opts is None: from vibeqc._vibeqc_core import DavidsonOptions dav_opts = DavidsonOptions() e_nuc = float(nuclear_repulsion_per_cell(system, lat_opts)) def diagonalize(F: np.ndarray) -> Tuple[np.ndarray, np.ndarray]: Fp = X.T @ F @ X Fp = 0.5 * (Fp + Fp.T) if use_dav and dav_opts is not None: from vibeqc._vibeqc_core import davidson_solve if dav_opts.n_eig == 0: dav_opts.n_eig = Fp.shape[0] if dav_opts.guess_vectors is not None: pass # already set from previous iteration dres = davidson_solve(Fp, dav_opts) if not dres.converged: raise RuntimeError( f"Davidson did not converge after {dres.n_iter} iters" ) eps, Cp = dres.eigenvalues, dres.eigenvectors dav_opts.guess_vectors = Cp else: eps, Cp = np.linalg.eigh(Fp) return X @ Cp, eps # Cache the iteration-invariant analytic-FT Hartree-J machinery once # (Bloch AO-pair FT, dense G-mesh, 4pi/G^2 kernel) so the PATOM # in-field step and the SCF loop reuse the same Ewald-J route. j_build = make_ewald_3d_gamma_j_builder( basis, system, omega=float(omega), lattice_opts=lat_opts, grid_shape=grid_shape_t, origin=origin, spacing_bohr=spacing_bohr, ) # ---- Initial guess via the unified engine. # Bug fix (v0.9.x): this Γ-only UHF path previously ignored # ``opts.initial_guess`` entirely and always used HCore. The # engine call respects the request; HCore reproduces the prior # behaviour bit-identically. guess = getattr(opts, "initial_guess", InitialGuess.HCORE) seed_guess = InitialGuess.SAD if guess == InitialGuess.PATOM else guess # Always seed the per-spin MO frame from Hcore -- used by the # quadratic-step tracking even when SAD provides the start density. C_alpha, eps_alpha = diagonalize(Hcore) C_beta, eps_beta = diagonalize(Hcore) split = initial_densities_open_shell( system.unit_cell_molecule(), basis, n_alpha, n_beta, seed_guess, is_periodic=True, periodic_system=system, lattice_opts=lat_opts, # ATOMSPIN: per-atom +1/-1/0 seed -> broken-symmetry g=0 density # (Bloch-sums to a broken-symmetry D(k)). Empty/None = symmetric. atomic_spins=getattr(opts, "atomic_spins", None) or None, # READ restart (Γ-only): prior per-spin g=0 cell densities. read_density_alpha=getattr(opts, "read_density_alpha", None), read_density_beta=getattr(opts, "read_density_beta", None), read_path=getattr(opts, "read_path", ""), ) if split is not None: D_alpha, D_beta = split else: # HCore fallback: per-spin densities from the leading # occupied columns of the Hcore-diagonalised MO frame. D_alpha = ( C_alpha[:, :n_alpha] @ C_alpha[:, :n_alpha].T if n_alpha > 0 else np.zeros_like(Hcore) ) D_beta = ( C_beta[:, :n_beta] @ C_beta[:, :n_beta].T if n_beta > 0 else np.zeros_like(Hcore) ) D_alpha = 0.5 * (D_alpha + D_alpha.T) D_beta = 0.5 * (D_beta + D_beta.T) if guess == InitialGuess.PATOM: plog.info("initial guess: PATOM (SAD + one periodic in-field step)") D_total = D_alpha + D_beta J = j_build(D_total) jk_alpha = build_jk_gamma_molecular_limit( basis, system, lat_opts, 2.0 * D_alpha, 0.0, ) jk_beta = build_jk_gamma_molecular_limit( basis, system, lat_opts, 2.0 * D_beta, 0.0, ) K_alpha = 0.5 * np.asarray(jk_alpha.K) K_beta = 0.5 * np.asarray(jk_beta.K) F_alpha_seed = 0.5 * ((Hcore + J - K_alpha) + (Hcore + J - K_alpha).T) F_beta_seed = 0.5 * ((Hcore + J - K_beta) + (Hcore + J - K_beta).T) C_alpha, eps_alpha = diagonalize(F_alpha_seed) C_beta, eps_beta = diagonalize(F_beta_seed) D_alpha = ( C_alpha[:, :n_alpha] @ C_alpha[:, :n_alpha].T if n_alpha > 0 else np.zeros_like(Hcore) ) D_beta = ( C_beta[:, :n_beta] @ C_beta[:, :n_beta].T if n_beta > 0 else np.zeros_like(Hcore) ) D_alpha = 0.5 * (D_alpha + D_alpha.T) D_beta = 0.5 * (D_beta + D_beta.T) D_alpha_prev = D_alpha.copy() D_beta_prev = D_beta.copy() damping = float(opts.damping) if not (0.0 <= damping < 1.0): raise ValueError( f"run_uhf_periodic_gamma_ewald3d: damping must be in [0, 1); got {damping}" ) damper: Optional[DynamicDamping] = None if bool(getattr(opts, "dynamic_damping", False)): damper = DynamicDamping( initial_alpha=damping, alpha_min=float(getattr(opts, "dynamic_damping_min", 0.0)), alpha_max=float(getattr(opts, "dynamic_damping_max", 0.95)), ) use_diis = bool(opts.use_diis) diis_start_iter = int(opts.diis_start_iter) accel: Optional[PeriodicSCFAccelerator] = ( PeriodicSCFAccelerator(opts) if use_diis else None ) level_shift = float(getattr(opts, "level_shift", 0.0)) # Explicit per-iteration schedule (unified with the molecular # drivers). Empty ⇒ the constant `level_shift` above; non-empty ⇒ # resolved per iteration by the shared C++ helper, applied per-spin. _ls_schedule = list(getattr(opts, "level_shift_schedule", None) or []) _ls_max_iter = int(opts.max_iter) # SPINLOCK PATTERN_HOLD: hold the seeded broken-symmetry occupied set by # maximum overlap (MOM) with the previous cycle for the first # ``spinlock_iterations`` cycles, then release -- protecting an ATOMSPIN # seed from collapsing to the symmetric solution. Mirrors the molecular # UHF path (cpp/src/uhf.cpp) and the periodic multi-k UKS driver. spinlock_mode = getattr(opts, "spinlock_mode", SpinlockMode.OFF) spinlock_iterations = int(getattr(opts, "spinlock_iterations", 0)) _pattern_hold = ( spinlock_mode == SpinlockMode.PATTERN_HOLD and spinlock_iterations > 0 ) # Phase C1c -- quadratic SCF fallback (per-spin Newton step). quadratic_fallback_iter = int(getattr(opts, "quadratic_fallback_iter", 0)) quadratic_fallback_shift = float(getattr(opts, "quadratic_fallback_shift", 0.1)) quadratic_fallback_max_step = float( getattr(opts, "quadratic_fallback_max_step", 0.1) ) # Track per-spin MO basis between iterations for the C1c step. C_alpha_prev_mo = C_alpha eps_alpha_prev_mo = eps_alpha C_beta_prev_mo = C_beta eps_beta_prev_mo = eps_beta # ---- SCF loop -------------------------------------------------------- scf_trace: List[SCFIteration] = [] result = PeriodicUHFEwaldResult( energy=0.0, e_electronic=0.0, e_nuclear=e_nuc, n_iter=0, converged=False, s_squared=0.0, s_squared_ideal=0.25 * (mult - 1) * (mult + 1), mo_energies_alpha=np.empty(0), mo_coeffs_alpha=np.empty((0, 0)), density_alpha=D_alpha.copy(), fock_alpha=np.empty((0, 0)), mo_energies_beta=np.empty(0), mo_coeffs_beta=np.empty((0, 0)), density_beta=D_beta.copy(), fock_beta=np.empty((0, 0)), overlap=S, scf_trace=scf_trace, omega=float(omega), grid_shape=grid_shape_t, ) E_prev = 0.0 plog.banner("SCF (UHF Gamma, EWALD_3D)") plog.info(" iter energy (Ha) dE ||[F,DS]|| DIIS") for iter_idx in range(1, int(opts.max_iter) + 1): if damper is not None: damping = damper.alpha # SPINLOCK PATTERN_HOLD: the accelerator is suspended (no history # recorded, no extrapolation, damping stays live) while the hold is # active. Fock extrapolation across held-window iterates steers the # SCF toward the symmetric attractor by continuous orbital rotation # -- a collapse the occupation-selecting MOM hold cannot see -- and # poisons the post-release history with out-of-basin iterates. The # history starts fresh at release. hold_active = _pattern_hold and iter_idx <= spinlock_iterations diis_active = ( use_diis and iter_idx >= diis_start_iter and not hold_active ) # Density damping per spin; skip when DIIS active. if iter_idx == 1 or damping == 0.0 or diis_active: D_alpha_used = D_alpha D_beta_used = D_beta else: D_alpha_used = damping * D_alpha_prev + (1.0 - damping) * D_alpha D_beta_used = damping * D_beta_prev + (1.0 - damping) * D_beta # Total density for J. Note the factor-2 convention: UHF # densities are one-particle (no factor 2); the J builder in # ``build_jk_gamma_molecular_limit`` expects the closed-shell # convention D_total = D_a + D_b (which gives the right Hartree # potential). ``build_j_ewald_3d`` follows the same convention. D_total = D_alpha_used + D_beta_used # Hartree J via composed Ewald-3D (cached analytic-FT machinery). J = j_build(D_total) # Per-spin K via the full-range real-space builder. The # closed-shell ``build_jk_gamma_molecular_limit`` returns # K(D) where D is the input density; for UHF we feed in # 2.D_a (the closed-shell-convention density that gives the # right K(D_a)) and divide the resulting K by 2 to recover # the per-spin exchange. # Equivalently: jk_a.K = K(2.D_a) = 2.K(D_a), so K(D_a) # = jk_a.K / 2. jk_alpha = build_jk_gamma_molecular_limit( basis, system, lat_opts, 2.0 * D_alpha_used, 0.0, ) jk_beta = build_jk_gamma_molecular_limit( basis, system, lat_opts, 2.0 * D_beta_used, 0.0, ) K_alpha = 0.5 * np.asarray(jk_alpha.K) K_beta = 0.5 * np.asarray(jk_beta.K) F_alpha = Hcore + J - K_alpha F_beta = Hcore + J - K_beta F_alpha = 0.5 * (F_alpha + F_alpha.T) F_beta = 0.5 * (F_beta + F_beta.T) # Per-cell electronic energy E_elec = ( 0.5 * float(np.einsum("ij,ij->", D_alpha_used + D_beta_used, Hcore)) + 0.5 * float(np.einsum("ij,ij->", D_alpha_used, F_alpha)) + 0.5 * float(np.einsum("ij,ij->", D_beta_used, F_beta)) ) # Madelung-leak correction (v0.6.1). E_madelung_fix = _madelung_energy_correction_for_lat( D_alpha_used + D_beta_used, S, system, lat_opts ) E_total = E_elec + e_nuc + E_madelung_fix # Orbital-gradient norms per spin. FDS_alpha = F_alpha @ D_alpha_used @ S FDS_beta = F_beta @ D_beta_used @ S grad_alpha = FDS_alpha - FDS_alpha.T grad_beta = FDS_beta - FDS_beta.T grad_norm = float( np.sqrt(np.linalg.norm(grad_alpha) ** 2 + np.linalg.norm(grad_beta) ** 2) ) dE = E_total - E_prev # Divergence detection (v0.6.2). check_scf_divergence( "run_uhf_periodic_gamma_ewald3d", iter_idx, E_total, grad_norm, dE, ) diis_sub = accel.subspace_size if accel is not None else 0 scf_trace.append( SCFIteration( iter=iter_idx, energy=float(E_total), delta_e=float(dE if iter_idx > 1 else 0.0), grad_norm=float(grad_norm), diis_subspace=diis_sub, ) ) plog.iteration( iter_idx, energy=float(E_total), dE=float(dE if iter_idx > 1 else 0.0), grad=float(grad_norm), diis=diis_sub, ) # Per-iter energy decomposition for cross-code parity comparison. D_total_iter = D_alpha_used + D_beta_used E_kin_iter = float(np.einsum("ij,ij->", D_total_iter, T)) E_ne_iter = float(np.einsum("ij,ij->", D_total_iter, V)) E_J_iter = 0.5 * float(np.einsum("ij,ij->", D_total_iter, J)) E_K_iter = -0.5 * float( np.einsum("ij,ij->", D_alpha_used, K_alpha) ) - 0.5 * float(np.einsum("ij,ij->", D_beta_used, K_beta)) plog.energy_decomposition( iter_idx, E_kin=E_kin_iter, E_ne=E_ne_iter, E_J=E_J_iter, E_K=E_K_iter, E_nuc=float(e_nuc), E_madelung=float(E_madelung_fix), ) converged = ( iter_idx > 1 and abs(dE) < float(opts.conv_tol_energy) and grad_norm < float(opts.conv_tol_grad) ) # Phase C1c gate -- per-spin Newton step replaces both DIIS and # diagonalisation when the fallback is active. in_quadratic_phase = ( quadratic_fallback_iter > 0 and iter_idx > quadratic_fallback_iter ) if in_quadratic_phase: from .quadratic_scf import quadratic_step C_alpha_new, eps_alpha_new = quadratic_step( F_alpha, C_alpha_prev_mo, eps_alpha_prev_mo, n_alpha, shift=quadratic_fallback_shift, max_step=quadratic_fallback_max_step, ) C_beta_new, eps_beta_new = quadratic_step( F_beta, C_beta_prev_mo, eps_beta_prev_mo, n_beta, shift=quadratic_fallback_shift, max_step=quadratic_fallback_max_step, ) else: # SCF-accelerator extrapolation. Every accelerator (DIIS # included) runs a single spin-coupled history -- one # coefficient set extrapolates both Focks; see the # _AcceleratorState note in periodic_scf_accelerators.py. # Skipped entirely (not even recorded) while the PATTERN_HOLD # window is active; see the hold_active note at the loop head. if accel is not None and not hold_active: F_alpha_ex, F_beta_ex = accel.extrapolate_uhf( F_alpha, F_beta, error_alpha=grad_alpha, error_beta=grad_beta, density_alpha=D_alpha_used, density_beta=D_beta_used, energy=E_total, mo_coeffs_alpha=C_alpha_prev_mo, mo_coeffs_beta=C_beta_prev_mo, mo_energies_alpha=eps_alpha_prev_mo, mo_energies_beta=eps_beta_prev_mo, n_alpha=n_alpha, n_beta=n_beta, ) if diis_active: F_alpha = F_alpha_ex F_beta = F_beta_ex # Saunders-Hillier level shift per spin. UHF densities are # one-particle (no factor 2), so the projector onto the # occupied subspace is S.D_s.S (not 1/2.S.D.S as for closed # shell) and the shift formula is # F_s_shift = F_s + b.S - b.(S.D_s.S) # The pre-update density D_s_used pins the shift to the # current iterate (matches the multi-k UHF and RHF/RKS # conventions). b = ( level_shift_at_iter( level_shift, 0, _ls_schedule, _ls_max_iter, iter_idx ) if _ls_schedule else level_shift ) # Shared Saunders-Hillier operator. D_alpha_used / D_beta_used are # *spin* densities (occupations in {0, 1}) and idempotent in the S # metric, so S·Dσ·S is the occupied projector and the weight is 1, # not the closed-shell ½. Returns F untouched when b == 0. if b == 0.0: # Skip the pybind conversion of S and Dσ entirely on an # unshifted cycle, which is the default and common case. F_alpha_diag = F_alpha F_beta_diag = F_beta else: F_alpha_diag = apply_level_shift( F_alpha, S, D_alpha_used, b, LevelShiftDensity.SPIN) F_beta_diag = apply_level_shift( F_beta, S, D_beta_used, b, LevelShiftDensity.SPIN) # Diagonalize + update. C_alpha_new, eps_alpha_new = diagonalize(F_alpha_diag) C_beta_new, eps_beta_new = diagonalize(F_beta_diag) # SPINLOCK PATTERN_HOLD: after both the normal and quadratic branches # produce C_*_new, reorder the occupied set by max overlap with the # previous cycle (MOM) for cycles 2..spinlock_iterations, then release. # iter 1 sets the pattern by aufbau (C_*_prev_mo is the prior cycle). if ( _pattern_hold and 1 < iter_idx <= spinlock_iterations and C_alpha_prev_mo.shape[1] >= n_alpha and C_beta_prev_mo.shape[1] >= n_beta ): if n_alpha > 0: C_alpha_new, eps_alpha_new = _mom_reorder( C_alpha_new, eps_alpha_new, S, C_alpha_prev_mo[:, :n_alpha], n_alpha ) if n_beta > 0: C_beta_new, eps_beta_new = _mom_reorder( C_beta_new, eps_beta_new, S, C_beta_prev_mo[:, :n_beta], n_beta ) C_alpha_prev_mo = C_alpha_new eps_alpha_prev_mo = eps_alpha_new C_beta_prev_mo = C_beta_new eps_beta_prev_mo = eps_beta_new D_alpha_prev = D_alpha_used D_beta_prev = D_beta_used if n_alpha > 0: D_alpha = C_alpha_new[:, :n_alpha] @ C_alpha_new[:, :n_alpha].T else: D_alpha = np.zeros_like(Hcore) if n_beta > 0: D_beta = C_beta_new[:, :n_beta] @ C_beta_new[:, :n_beta].T else: D_beta = np.zeros_like(Hcore) D_alpha = 0.5 * (D_alpha + D_alpha.T) D_beta = 0.5 * (D_beta + D_beta.T) # Stash latest state on the result for inspection even if not # converged. result.energy = E_total result.e_electronic = E_elec result.n_iter = iter_idx result.mo_energies_alpha = eps_alpha_new result.mo_coeffs_alpha = C_alpha_new result.density_alpha = D_alpha_used result.fock_alpha = F_alpha result.mo_energies_beta = eps_beta_new result.mo_coeffs_beta = C_beta_new result.density_beta = D_beta_used result.fock_beta = F_beta if damper is not None: damper.update(E_total) E_prev = E_total if converged: # Final consistency pass on the fresh D's (matches RHF # driver convention; reports the un-shifted physical Fock # / MOs). D_total_f = D_alpha + D_beta J_f = j_build(D_total_f) jk_a_f = build_jk_gamma_molecular_limit( basis, system, lat_opts, 2.0 * D_alpha, 0.0, ) jk_b_f = build_jk_gamma_molecular_limit( basis, system, lat_opts, 2.0 * D_beta, 0.0, ) F_alpha_f = Hcore + J_f - 0.5 * np.asarray(jk_a_f.K) F_beta_f = Hcore + J_f - 0.5 * np.asarray(jk_b_f.K) F_alpha_f = 0.5 * (F_alpha_f + F_alpha_f.T) F_beta_f = 0.5 * (F_beta_f + F_beta_f.T) C_a_f, eps_a_f = diagonalize(F_alpha_f) C_b_f, eps_b_f = diagonalize(F_beta_f) # If the SCF converged while the PATTERN_HOLD window was still # active, the converged state is the MOM-held occupied set, # which need not be aufbau in its own Fock. Re-select by max # overlap with the held pattern so the reported energy / <S^2> # / MOs describe the state the SCF actually converged to (an # aufbau slice here would silently swap to a different state). if ( _pattern_hold and iter_idx <= spinlock_iterations and C_alpha_prev_mo.shape[1] >= n_alpha and C_beta_prev_mo.shape[1] >= n_beta ): if n_alpha > 0: C_a_f, eps_a_f = _mom_reorder( C_a_f, eps_a_f, S, C_alpha_prev_mo[:, :n_alpha], n_alpha ) if n_beta > 0: C_b_f, eps_b_f = _mom_reorder( C_b_f, eps_b_f, S, C_beta_prev_mo[:, :n_beta], n_beta ) E_elec_f = ( 0.5 * float(np.einsum("ij,ij->", D_alpha + D_beta, Hcore)) + 0.5 * float(np.einsum("ij,ij->", D_alpha, F_alpha_f)) + 0.5 * float(np.einsum("ij,ij->", D_beta, F_beta_f)) ) E_madelung_fix_f = _madelung_energy_correction_for_lat( D_alpha + D_beta, S, system, lat_opts ) result.energy = E_elec_f + e_nuc + E_madelung_fix_f result.e_electronic = E_elec_f result.mo_energies_alpha = eps_a_f result.mo_coeffs_alpha = C_a_f result.density_alpha = D_alpha result.fock_alpha = F_alpha_f result.mo_energies_beta = eps_b_f result.mo_coeffs_beta = C_b_f result.density_beta = D_beta result.fock_beta = F_beta_f result.s_squared = _spin_squared( n_alpha, n_beta, C_a_f, C_b_f, S, ) result.converged = True plog.converged( n_iter=result.n_iter, energy=result.energy, converged=True, ) return result # Did not converge -- populate <S^2> on the final iterate anyway. result.s_squared = _spin_squared( n_alpha, n_beta, result.mo_coeffs_alpha, result.mo_coeffs_beta, S, ) plog.converged( n_iter=result.n_iter, energy=result.energy, converged=False, ) return result