Source code for vibeqc.periodic_rhf_ewald

"""Phase 12e-c-4b: Γ-point periodic RHF SCF driver using the
composed Ewald-3D Coulomb dispatch.

Replaces the ``DIRECT_TRUNCATED`` J build used by the existing
C++ ``run_rhf_periodic_gamma`` driver with the ω-invariant composed
J from :func:`vibeqc.build_j_ewald_3d` (Phase 12e-c-4a). Exchange K
stays on the full-range real-space builder
(:func:`vibeqc.build_jk_gamma_molecular_limit` with ω = 0) because
K inherits its locality from the density matrix rather than the
Coulomb operator — standard periodic-HF practice.

What this driver does
---------------------

At every SCF iteration:

    F  =  H_core  +  J_ewald(ω, D)  -  ½ K(D)

with H_core = T + V (both periodic one-electron integrals, Bloch-
summed at Γ), S the Γ-point overlap, and canonical-orthogonalisation
replacing symmetric orthogonalisation so that linearly-dependent
bases converge rather than crash. Convergence check is identical to
the C++ driver: ``|ΔE| < conv_tol_energy`` and
``‖F D S − S D F‖_F < conv_tol_grad``.

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

* **Γ-only** — single k-point at (0, 0, 0). Multi-k EWALD_3D is a
  follow-up (12e-c-4c) that needs the long-range J FFT generalised
  across Bloch-summed densities.
* **Molecular-limit cells** — the cell has to be big enough that
  the density doesn't overlap its periodic image through the box
  boundary. For real bulk cells (tight unit cells with cell-to-cell
  density overlap) the FFT density grid needs a multi-cell sum, also
  deferred to 12e-c-4c.
* **DIIS supported** via ``options.use_diis = True`` (the default on
  :class:`PeriodicRHFOptions`). Pure-Python Pulay-DIIS matching the
  C++ molecular implementation (same subspace size, same error-vector
  ``F D S − S D F``); typically halves iteration counts on
  molecular-limit SCF cases vs plain damping.

Validation targets
------------------

* **ω-invariance**: total SCF energy must be the same (to ~µHa) for
  ω ∈ [0.3, 2.0] when applied to the same system.
* **Cross-check against DIRECT_TRUNCATED at molecular limit**: for a
  molecule in a ~30 bohr vacuum box, the EWALD_3D and DIRECT_TRUNCATED
  drivers must agree to within Makov–Payne finite-box corrections
  (~O(1/L³), sub-mHa at this box size).

These are exercised in ``tests/test_periodic_rhf_ewald.py``.
"""

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 (
    DIIS,
    BasisSet,
    CoulombMethod,
    InitialGuess,
    LatticeSumOptions,
    PeriodicRHFOptions,
    PeriodicSystem,
    SCFIteration,
    bloch_sum,
    build_jk_gamma_molecular_limit,
    compute_kinetic_lattice,
    compute_nuclear_lattice,
    compute_overlap_lattice,
    nuclear_repulsion_per_cell,
)
from .ewald_composed import build_j_ewald_3d
from .ewald_j import auto_grid
from .guess import initial_density_closed_shell
from .linear_dependence import scf_preflight_overlap_check
from .options_dump import dump_active_settings
from .periodic_scf_accelerators import DynamicDamping, PeriodicSCFAccelerator
from .progress import ProgressLogger, resolve_progress
from .scf_divergence import check_scf_divergence

__all__ = [
    "PeriodicRHFEwaldResult",
    "run_rhf_periodic_gamma_ewald3d",
]


[docs] @dataclass class PeriodicRHFEwaldResult: """Result of :func:`run_rhf_periodic_gamma_ewald3d`. Structurally parallel to :class:`PeriodicRHFResult` but carries Ewald-specific bookkeeping (``omega``, ``grid_shape``) so callers can reproduce the calculation or sweep parameters. All matrices are at Γ-point (no Bloch sums). """ energy: float e_electronic: float e_nuclear: float n_iter: int converged: bool mo_energies: np.ndarray mo_coeffs: np.ndarray density: np.ndarray fock: np.ndarray overlap: np.ndarray scf_trace: List[SCFIteration] = field(default_factory=list) # Ewald-specific: omega: float = 0.0 grid_shape: Tuple[int, int, int] = (0, 0, 0)
# Pulay DIIS — re-export the C++ binding ``vibeqc::DIIS`` under the legacy # internal name ``_PulayDIIS``. Until 2026-05-10 this module shipped its # own pure-Python Pulay-DIIS class; the C++ binding is the canonical # implementation and the duplicate Python copy is redundant under § 10 # ("shared infrastructure between molecular and periodic"). The binding # uses ``Eigen::FullPivLU`` for the Pulay linear solve, which is more # numerically stable than the previous ``np.linalg.solve`` + # coefficient-blow-up retry; behaviour is equivalent at the SCF fixed # point. The legacy name is preserved so call sites elsewhere in # ``python/vibeqc/`` need no churn. _PulayDIIS = DIIS def _reject_unsupported_python_accelerator(opts, driver: str) -> None: """Reject non-DIIS accelerator + dynamic-damping in Python periodic backends. Background. ``PeriodicRHFOptions`` / ``PeriodicSCFOptions`` / ``PeriodicKSOptions`` expose ``scf_accelerator`` (DIIS / KDIIS / EDIIS / EDIIS_DIIS / ADIIS) and ``dynamic_damping`` to match the molecular options surface. The C++ direct-truncated periodic kernels (``periodic_rhf.cpp``, ``periodic_scf.cpp``) consume these fields, but the Python periodic backends (Ewald / GDF / BIPOLE) only implement Pulay DIIS + static damping today. Silently ignoring a user's ``scf_accelerator = EDIIS_DIIS`` request was the bug; this helper raises a clear ``NotImplementedError`` so users get a pointer to the C++ direct path instead of an unexpected DIIS run. """ from ._vibeqc_core import SCFAccelerator accel = getattr(opts, "scf_accelerator", SCFAccelerator.DIIS) if accel != SCFAccelerator.DIIS: name = getattr(accel, "name", str(accel)) raise NotImplementedError( f"{driver}: scf_accelerator={name} is not implemented on this " "Python periodic backend. Only SCFAccelerator.DIIS is wired " "today. The C++ direct-truncated periodic path " "(CoulombMethod.DIRECT_TRUNCATED) supports KDIIS / EDIIS / " "EDIIS_DIIS / ADIIS — switch to that route if you need a " "non-DIIS accelerator. See docs/user_guide/scf_convergence.md " "for the current support matrix." ) if bool(getattr(opts, "dynamic_damping", False)): raise NotImplementedError( f"{driver}: opts.dynamic_damping = True is not implemented on " "this Python periodic backend. Only static opts.damping is " "honoured today. The C++ direct-truncated periodic path " "supports dynamic damping (Zerner-Hehenberger 1979) — switch " "to CoulombMethod.DIRECT_TRUNCATED if you need it." ) def _canonical_orthogonalizer( S: np.ndarray, threshold: float = 1e-7, *, normalize_diag_first: bool = True, ) -> Tuple[np.ndarray, int]: """Return the (n_bf, n_kept) rectangular orthogonaliser X such that ``X^T S X ≈ I_{n_kept}``, with any S-eigenvalue below ``threshold`` dropped (Löwdin canonical orthogonalisation). Parameters ---------- S ``(n, n)`` symmetric overlap matrix. Real-valued; for the complex Hermitian (multi-k) case use :func:`_canonical_orthogonalizer_complex` (in ``periodic_rhf_multi_k_ewald``). threshold Eigenvalues below this are projected out. ``1e-7`` matches the PySCF default. normalize_diag_first When ``True`` (default, PySCF-aligned), pre-conditioning scales rows / columns of S by ``1 / sqrt(diag(S))`` before the eigendecomposition; the eigenvalues then live on a unit-diagonal matrix where the threshold has its intended meaning. After diagonalisation the normalisation is undone so X still satisfies ``X^T S X = I``. Mirrors PySCF's ``pyscf.scf.addons.canonical_orth_`` (and the comment "Ensure the basis functions are normalized — symmetry-adapted ones are not!"). When ``False`` (vibe-qc legacy ≤ v0.6), eigh runs on the bare S. Cheaper, but the threshold is harder to interpret on bases whose diagonal entries vary by orders of magnitude. Kept available for cross-checking against pre-v0.7 numbers. Returns ------- X : np.ndarray of shape ``(n_bf, n_kept)`` Rectangular orthogonaliser. n_kept : int Number of S-eigenvectors retained. References ---------- Löwdin, P.-O. *Adv. Quantum Chem.* **5**, 185 (1970) — canonical orthogonalisation in molecular quantum chemistry. Lykos, P. G., Schmeising, H. N. *J. Chem. Phys.* **35**, 288 (1961) — early analysis of the AO overlap eigenvalue spectrum. """ if normalize_diag_first: # PySCF-aligned path. Diagonal scaling: D = diag(1/sqrt(S_ii)). # Snorm = D · S · D has unit diagonal so the threshold operates # on a "shape-only" overlap and is robust to per-AO scale. d = np.asarray(S).diagonal() # Guard against zero or negative diagonals (shouldn't happen # for a real basis, but the preflight check might let through # numerical near-zero entries). if np.any(d <= 0): # Fall back to raw-S path for the affected rows; canonical # orth on the remainder. normalize_diag_first = False else: scale = 1.0 / np.sqrt(d) Snorm = S * np.outer(scale, scale) eigvals, eigvecs = np.linalg.eigh(Snorm) mask = eigvals >= threshold n_kept = int(mask.sum()) if n_kept == 0: raise RuntimeError( "run_rhf_periodic_gamma_ewald3d: no overlap eigenvalue " f"above threshold {threshold:.1e} (after sqrt-diag " "normalisation); basis is fully linearly dependent" ) kept_vals = eigvals[mask] kept_vecs = eigvecs[:, mask] X_norm = kept_vecs / np.sqrt(kept_vals) # Plug normalisation back in: X = D · X_norm so X^T S X = I. X = scale[:, None] * X_norm return X, n_kept # Legacy raw-S path — kept for cross-checking pre-v0.7 numbers. eigvals, eigvecs = np.linalg.eigh(S) mask = eigvals >= threshold n_kept = int(mask.sum()) if n_kept == 0: raise RuntimeError( "run_rhf_periodic_gamma_ewald3d: no overlap eigenvalue " f"above threshold {threshold:.1e}; basis is fully " "linearly dependent" ) kept_vals = eigvals[mask] kept_vecs = eigvecs[:, mask] X = kept_vecs / np.sqrt(kept_vals) return X, n_kept
[docs] def run_rhf_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, progress: Union[bool, ProgressLogger, None] = None, verbose: Optional[int] = None, ) -> PeriodicRHFEwaldResult: """Γ-point closed-shell periodic RHF SCF with Ewald-3D Coulomb. The Coulomb matrix is built via the composed Ewald split ``J_ewald(ω, D) = J_SR(ω) + J_LR(ω)`` from :func:`build_j_ewald_3d`. Exchange K uses the full-range real-space builder (``build_jk_gamma_molecular_limit`` at ω = 0); in periodic HF, K's real-space decay comes from the density matrix, not the Coulomb operator, so an Ewald split on K is neither needed nor used by standard periodic codes. Parameters ---------- system :class:`PeriodicSystem`. basis AO basis for the unit cell. options Optional :class:`PeriodicRHFOptions` controlling max_iter, damping, convergence tolerances. ``use_diis`` is currently ignored — plain damping only; DIIS is a planned follow-up. omega Ewald splitting parameter. Result is ω-independent at convergence to ~ µHa across ω ∈ [0.3, 2.0]. grid_shape, origin, spacing_bohr FFT-Poisson grid controls forwarded to :func:`build_j_ewald_3d`. linear_dep_threshold Overlap-eigenvalue threshold below which AO directions are projected out of the SCF. Default 1e-7 matches the molecular SCF drivers. Returns ------- :class:`PeriodicRHFEwaldResult`. """ opts = options if options is not None else PeriodicRHFOptions() lat_opts = opts.lattice_opts # ---- Force EWALD_3D gauge ---- if lat_opts.coulomb_method != CoulombMethod.EWALD_3D: lat_opts.coulomb_method = CoulombMethod.EWALD_3D if lat_opts.cutoff_bohr < 18.0 and system.dim == 3: V_cell = float(abs(np.linalg.det(np.asarray(system.lattice)))) if V_cell < 1000.0: lat_opts.cutoff_bohr = max(lat_opts.cutoff_bohr, 18.0) lat_opts.nuclear_cutoff_bohr = max(lat_opts.nuclear_cutoff_bohr, 25.0) plog = resolve_progress(progress, verbose=verbose) # ---- Ewald splitting parameter ω — must match nuclear α ---- # When not specified explicitly, use CRYSTAL's default α = 2.8/V^{1/3} # (matching the BIPOLE driver and CRYSTAL's MADEL2). The old # geometry-based formula √(−log tol)/cutoff had no relationship to # the nuclear Ewald α and broke gauge consistency. _ewald_tol = getattr(opts, "ewald_tolerance", 1e-12) _cutoff = getattr(opts, "ewald_cutoff_bohr", lat_opts.nuclear_cutoff_bohr) if omega <= 0.0: # omega not explicitly set by caller; check options, then CRYSTAL default. _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"RHF 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 — every overridable knob gets printed so # the SCF log self-documents its inputs. See vibeqc.options_dump. 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": True, }, ), ], ) if plog.level >= 5: from .scf_log import format_basis_summary plog.write_raw(format_basis_summary(basis)) # Closed-shell check n_elec = system.n_electrons() if n_elec % 2 != 0: raise ValueError( "run_rhf_periodic_gamma_ewald3d: closed-shell RHF " f"requires even electron count; got {n_elec}" ) if system.multiplicity != 1: raise ValueError( "run_rhf_periodic_gamma_ewald3d: closed-shell RHF requires " f"multiplicity=1; got {system.multiplicity}" ) n_occ = n_elec // 2 # ---- Auto-optimise lattice truncation (default ON) ------------------- # See periodic_rks_ewald.py for the rationale. 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.warning( "auto_optimize_truncation did not converge; SCF will run " "with the last evaluated settings. Check the report and " "consider vq.disambiguate_critical_overlap and / or " "vq.make_basis(..., exp_to_discard=...)." ) 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 # Symmetrize against tiny imaginary drift from the Bloch sum. S = 0.5 * (S + S.T) Hcore = 0.5 * (Hcore + Hcore.T) # ---- Linear-dependence preflight on S(Γ) ------------------------------ # PySCF only logs this at DEBUG verbosity; we surface it at INFO # because near-singular / non-PSD S in periodic SCF is much more # common than in molecular SCF (especially after the v0.7 periodic # AO image-summing) and silently truncating the basis via canonical # orthogonalisation can give plausible-but-wrong total energies. # Negative S eigenvalues (severity="critical") abort by default. scf_preflight_overlap_check( S, plog=plog, label="S(Γ)", basis=basis, ) # ---- Canonical orthogonalisation -------------------------------------- X, n_kept = _canonical_orthogonalizer( S, linear_dep_threshold, normalize_diag_first=canonical_orth_normalize_diag_first, ) if n_occ > n_kept: raise RuntimeError( "run_rhf_periodic_gamma_ewald3d: canonical orthogonalisation " f"dropped too many directions (n_occ = {n_occ}, " f"n_kept = {n_kept}); loosen linear_dep_threshold or pick " "a less redundant basis." ) # ---- Nuclear repulsion per cell --------------------------------------- e_nuc = nuclear_repulsion_per_cell(system, lat_opts) # ---- Madelung-leak correction (v0.6.1 fix) --------------------------- # See vibeqc.madelung.madelung_energy_correction for the rationale. from .madelung import madelung_energy_correction # ---- Initial guess via the unified engine. # SAD is the right default for ionic / covalent insulators (NaCl, # MgO, Si) where Hcore underbinds the deep core states so badly # that DIIS can amplify the swing into +30 000 / −16 000 Ha # territory before recovering. def diagonalise(F): Fp = X.T @ F @ X Fp = 0.5 * (Fp + Fp.T) eps, Cp = np.linalg.eigh(Fp) C = X @ Cp return C, eps guess = getattr(opts, "initial_guess", InitialGuess.HCORE) D_engine = initial_density_closed_shell( system.unit_cell_molecule(), basis, n_occ, guess, is_periodic=True, ) if D_engine is not None: plog.info(f"initial guess: {guess.name} (density via GuessEngine)") D = D_engine C0, eps0 = diagonalise(Hcore) # MO trace for log symmetry else: plog.info(f"initial guess: {guess.name} (Hcore-diagonalise)") C0, eps0 = diagonalise(Hcore) D = 2.0 * C0[:, :n_occ] @ C0[:, :n_occ].T D = 0.5 * (D + D.T) D_prev = D.copy() damping = float(opts.damping) if not (0.0 <= damping < 1.0): raise ValueError( f"run_rhf_periodic_gamma_ewald3d: damping must be in [0, 1); got {damping}" ) # Dynamic damping (Zerner-Hehenberger 1979) — adjusts α between # ``[dynamic_damping_min, dynamic_damping_max]`` based on the # iteration-to-iteration energy decrease. Static damping is the # ``opts.dynamic_damping = False`` special case. 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)), ) # SCF accelerator — dispatches on opts.scf_accelerator. DIIS / # KDIIS / EDIIS / EDIIS_DIIS / ADIIS all share the # ``diis_start_iter`` activation gate; damping is suppressed once # the accelerator is active because damping + Fock-extrapolation # mixes badly. use_diis = bool(opts.use_diis) diis_start_iter = int(opts.diis_start_iter) accel: Optional[PeriodicSCFAccelerator] = ( PeriodicSCFAccelerator(opts) if use_diis else None ) # Saunders-Hillier level shift — opts.level_shift is 0.0 by default # (no shift). ``b > 0`` raises virtual MO eigenvalues by ``b`` # before each diagonalisation, suppressing occupied / virtual swaps # on small-gap insulators. level_shift = float(getattr(opts, "level_shift", 0.0)) # CRYSTAL-style FMIXING: linear blend with the previous iteration's # Fock matrix. ``fock_mixing_value`` is the weight of the previous # Fock; 0.0 disables the mix, 0.3 reproduces CRYSTAL14's default # FMIXING 30, 0.5–0.7 helps tough ionic SCF (LiH, NaCl, MgO # primitive). Applied AFTER DIIS extrapolation and BEFORE the # level-shift / diagonalisation step — same order as the gdf # driver (see periodic_rhf_gdf.py). fock_mixing_value = float(getattr(opts, "fock_mixing", 0.0)) if not (0.0 <= fock_mixing_value < 1.0): raise ValueError( "run_rhf_periodic_gamma_ewald3d: fock_mixing must be in [0, 1); " f"got {fock_mixing_value}" ) if fock_mixing_value != 0.0: plog.info( "fock mixing: CRYSTAL FMIXING " f"{100.0 * fock_mixing_value:.1f}% " "(previous Fock matrix weight)" ) F_prev_mixed: Optional[np.ndarray] = None # Phase C1c — second-order ("quadratic") SCF fallback. Inactive # by default. When opts.quadratic_fallback_iter > 0 and the SCF # has run that many iterations without converging, switch from # diagonalizing F to a Newton step in MO space (κ_{ai} = # −F_{ai}^MO / (ε_a − ε_i + λ); C_new = C_prev · exp(κ)). DIIS # and level shift are skipped during the quadratic phase. 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) ) # ---- SCF loop --------------------------------------------------------- plog.banner("SCF (RHF Gamma, EWALD_3D)") plog.info(" iter energy (Ha) dE ||[F,DS]|| DIIS") scf_trace: List[SCFIteration] = [] result = PeriodicRHFEwaldResult( energy=0.0, e_electronic=0.0, e_nuclear=float(e_nuc), n_iter=0, converged=False, mo_energies=np.empty(0), mo_coeffs=np.empty((0, 0)), density=D.copy(), fock=np.empty((0, 0)), overlap=S, scf_trace=scf_trace, omega=float(omega), grid_shape=grid_shape_t, ) E_prev = 0.0 F_final = Hcore.copy() C_final = C0 eps_final = eps0 for iter_idx in range(1, int(opts.max_iter) + 1): # Dynamic-damping read: ``damper.alpha`` carries the α # selected by the previous iteration's energy step. On iter 1 # this equals the initial α (the previous-iteration energy # has not been seen yet). if damper is not None: damping = damper.alpha diis_active = use_diis and iter_idx >= diis_start_iter # When DIIS is driving convergence, density damping is # unnecessary and actually slows things down by blurring # the density DIIS is trying to iterate. Match the C++ # driver and skip damping once DIIS is active. D_used = ( D if (iter_idx == 1 or damping == 0.0 or diis_active) else damping * D_prev + (1.0 - damping) * D ) # J via composed Ewald-3D (SR erfc + LR FFT) J = build_j_ewald_3d( basis, system, D_used, omega=float(omega), lattice_opts=lat_opts, grid_shape=grid_shape_t, origin=origin, spacing_bohr=spacing_bohr, ) # K via full-range real-space builder (omega=0). At Γ-only with # N_k = 1 this is the inverse-Bloch convention P(g) = P_Γ·δ_{g,0} # — density localized to the home cell, cross-cell P(h≠0) set # to zero (no information available without multi-k sampling). # The 2026-05-14 handover originally framed this as a builder # bug; the cutoff-convergence sweep on 2026-05-15 showed the # alternative "homogeneous-D" build diverges and the # molecular-limit kernel is in fact the correct Γ-only build. # Tight-ionic-crystal accuracy needs multi-k sampling. jk_full = build_jk_gamma_molecular_limit( basis, system, lat_opts, D_used, 0.0, ) K = np.asarray(jk_full.K) F = Hcore + J - 0.5 * K F = 0.5 * (F + F.T) # Closed-shell electronic energy: ½ tr[D (Hcore + F)] E_elec = 0.5 * float(np.einsum("ij,ij->", D_used, Hcore + F)) # Madelung-leak correction (v0.6.1) — disabled for EWALD_3D in # v0.7.0 since V_ne now dispatches to the Ewald path that's # gauge-aligned with J. See ``madelung_energy_correction_for_lat`` # docstring + LiH dense-ionic decomposition in # ``examples/debug/lih_energy_decomposition.py``. if lat_opts.coulomb_method == CoulombMethod.EWALD_3D: E_madelung_fix = 0.0 else: E_madelung_fix = madelung_energy_correction( D_used, S, system, nuclear_uses_ewald=False, ) E_total = E_elec + float(e_nuc) + E_madelung_fix # Orbital-gradient norm FDS = F @ D_used @ S grad = FDS - FDS.T grad_norm = float(np.linalg.norm(grad)) dE = E_total - E_prev # Divergence detection (v0.6.2 generalisation of v0.5.6 hook). check_scf_divergence( "run_rhf_periodic_gamma_ewald3d", iter_idx, E_total, grad_norm, dE, ) converged = ( iter_idx > 1 and abs(dE) < float(opts.conv_tol_energy) and grad_norm < float(opts.conv_tol_grad) ) 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=(accel.subspace_size if accel is not None else 0), ) ) plog.iteration( iter_idx, energy=float(E_total), dE=float(dE if iter_idx > 1 else 0.0), grad=float(grad_norm), diis=(accel.subspace_size if accel is not None else 0), ) # Per-iter energy decomposition for cross-code parity comparison. E_kin_iter = float(np.einsum("ij,ij->", D_used, T)) E_ne_iter = float(np.einsum("ij,ij->", D_used, V)) E_J_iter = 0.5 * float(np.einsum("ij,ij->", D_used, J)) E_K_iter = -0.5 * 0.5 * float(np.einsum("ij,ij->", D_used, K)) 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), ) # Phase C1c gate: when the quadratic fallback is enabled and # we've passed its activation iteration, take a Newton step # in MO space instead of diagonalizing. DIIS extrapolation # and level shift are both skipped — the Newton step is its # own update mechanism and mixing with DIIS undoes the # trust-region cap. in_quadratic_phase = ( quadratic_fallback_iter > 0 and iter_idx > quadratic_fallback_iter ) if in_quadratic_phase: from .quadratic_scf import quadratic_step C_new, eps_new = quadratic_step( F, C_final, eps_final, n_occ, shift=quadratic_fallback_shift, max_step=quadratic_fallback_max_step, ) else: # SCF-accelerator extrapolation: feed the current # iterate's matrices into the active accelerator and # replace F with the extrapolated Fock for the # diagonalisation step. The convergence check above uses # the *pre-extrapolation* F so dE and grad_norm still # reflect the raw SCF progress. Pass C_final / eps_final # from the previous iteration (the C ε pair that produced # the current density D from which F was built) — these # are the canonical-MO inputs KDIIS needs to project the # orbital-rotation gradient. if accel is not None: F_ex = accel.extrapolate_rhf( F, error=grad, density=D_used, energy=E_total, mo_coeffs=C_final, mo_energies=eps_final, n_occ=n_occ, ) if diis_active: F = F_ex # CRYSTAL-style FMIXING — applied after DIIS, before level- # shift / diag. Smooths Fock-matrix oscillations on tough # ionic SCF where the proper periodic J/K has two stationary # points (physical, over-bound). Matches the gdf driver's # FMIXING order. if fock_mixing_value != 0.0: if F_prev_mixed is not None: F_mixed = ( 1.0 - fock_mixing_value ) * F + fock_mixing_value * F_prev_mixed F = 0.5 * (F_mixed + F_mixed.T) F_prev_mixed = F.copy() # Apply Saunders-Hillier level shift to F before diagonalisation: # F_shifted = F + b · S − (b/2) · S · D · S # Raises virtual eigenvalues by b in the new MO basis. The shift # leaves the SCF fixed point unchanged but suppresses occupied / # virtual mixing during iteration. We use the *un-extrapolated* # density D_used (pre-update) so the shift respects the current # iterate, not the next. if level_shift != 0.0: F_diag = F + level_shift * S - (level_shift / 2.0) * (S @ D_used @ S) F_diag = 0.5 * (F_diag + F_diag.T) else: F_diag = F C_new, eps_new = diagonalise(F_diag) D_prev = D_used D = 2.0 * C_new[:, :n_occ] @ C_new[:, :n_occ].T D = 0.5 * (D + D.T) C_final = C_new eps_final = eps_new F_final = F # Dynamic-damping update — feed the just-computed E_total so # the *next* iteration reads an α adapted to this iteration's # energy step. No-op when ``damper is None`` (static damping). if damper is not None: damper.update(E_total) E_prev = E_total # Stash the per-iter state on the result for inspection even # if we don't converge. result.energy = E_total result.e_electronic = E_elec result.n_iter = iter_idx result.mo_energies = eps_new result.mo_coeffs = C_new result.density = D_used result.fock = F if converged: # Final self-consistency pass on the fresh D (same # convention as the C++ driver). J_f = build_j_ewald_3d( basis, system, D, omega=float(omega), lattice_opts=lat_opts, grid_shape=grid_shape_t, origin=origin, spacing_bohr=spacing_bohr, ) jk_f = build_jk_gamma_molecular_limit( basis, system, lat_opts, D, 0.0, ) K_f = np.asarray(jk_f.K) F_f = Hcore + J_f - 0.5 * K_f F_f = 0.5 * (F_f + F_f.T) C_f, eps_f = diagonalise(F_f) E_elec_f = 0.5 * float(np.einsum("ij,ij->", D, Hcore + F_f)) if lat_opts.coulomb_method == CoulombMethod.EWALD_3D: E_madelung_fix_f = 0.0 else: E_madelung_fix_f = madelung_energy_correction( D, S, system, nuclear_uses_ewald=False, ) result.energy = E_elec_f + float(e_nuc) + E_madelung_fix_f result.e_electronic = E_elec_f result.mo_energies = eps_f result.mo_coeffs = C_f result.density = D result.fock = F_f result.converged = True plog.converged( n_iter=result.n_iter, energy=result.energy, converged=True, ) return result result.converged = False plog.converged( n_iter=result.n_iter, energy=result.energy, converged=False, ) return result