DFT Production-Core Milestone

The DFT package contains two intentionally different surfaces. The legacy DFTSystem/run_scf surface supplies tiny Γ-point teaching, dense-reference, spin, occupation, finite-difference stress, and restart diagnostics. The periodic PeriodicDFTSystem/run_periodic_scf surface supplies the materials-workload path: PBE-PW92, reciprocal-space GTH and scalar norm-conserving UPF operators, Monkhorst-Pack integration, block-Davidson/Rayleigh-Ritz solves, fixed or Fermi-Dirac occupations, frozen-density band paths, and periodic forces across ordinary fixed full-rank cells. Analytic periodic stress and a restartable variable-cell surface are verified for one bounded 2H-Silicon workflow.

The periodic implementation has verified results for specific workloads, but is not broadly chemically certified. Capability claims are tied to the material-validation summary.

Nonlocal Pseudopotentials

The legacy UPF and GTH metadata path converts projectors into normalized real-space separable forms. The operator applies:

V̂_NL ψ = Σᵢ |βᵢ⟩ Dᵢ ⟨βᵢ|ψ⟩

The UPF parser preserves radial quadrature and the complete PP_DIJ matrix after Ry-to-Hartree conversion, while the legacy proof operator still consumes only its diagonal. The periodic scalar norm-conserving path implements the source-matched local transform, compact nonlocal SCF and bands, analytic fixed-cell forces, and content-bound checkpoints. Ultrasoft augmentation, PAW, spin-orbit terms, nonlinear core correction, and UPF analytic stress fail closed. The production periodic GTH path retains its separate analytic radial operators.

SCF applies nonlocal projectors by default when available. SCFConfig(apply_nonlocal=False) keeps the old local-only path available for debugging and comparison.

Solvers

Dense diagonalization remains the tiny-grid reference for the legacy path. Periodic SCF and bands use the MLX-native block-Davidson/Rayleigh-Ritz solver without building the full plane-wave Hamiltonian. Diagnostics expose residuals, orthonormality error, subspace work, and convergence metadata.

Periodic SCF reports one device-inclusive effective_potential timing for the independent Hartree and exchange-correlation branches. They share a single MLX materialization boundary so phase accounting does not serialize the runtime.

Adaptive periodic SCF uses paired subspace residuals only while its requested eigensolver tolerance is looser than the final Davidson tolerance. It restores direct-operator residual validation at the final tolerance, and SCF convergence requires that directly validated result. Fixed-tolerance and standalone eigensolves always retain direct validation.

Spin, Occupations, k-Points, And Bands

Time-reversal k-point ownership is SCF-safe because paired states contribute the same real-space density. General point-group reduction retains every full- mesh member, original weight, and reciprocal operation. Periodic SCF applies exact device-resident FFT-grid permutations to reconstruct scalar and collinear- spin densities before mixing. The caller must still establish that the supplied operations are symmetries of the cell, ions, and Hamiltonian. Incompatible grid operations fail closed. Forces and stress reject point-group-reduced results until their vector and tensor orbit reconstruction is implemented.

The new spin layer is collinear only:

• unpolarized: one total density ρ(r).
• polarized: separate ρ↑(r) and ρ↓(r) diagnostics.

The legacy layer exposes fixed and Fermi-Dirac occupation diagnostics. The periodic layer supports two complete occupation paths:

• The default fixed path computes exactly N/2 doubly occupied bands. It retains the cached-density fast path used by the verified insulating workloads.
• PeriodicFermiDiracSmearing(width_hartree=...) resolves one global chemical potential over the weighted k-point mesh. The caller supplies enough computed bands that 2 * n_bands > electron_count.

The smeared density and band energy use the resolved occupation of every band, not a post-hoc scalar density correction. SCF convergence follows the variational electronic free energy F = E - (k_B T) S; PeriodicSCFResult separately reports internal energy, chemical potential, dimensionless electronic entropy, and smearing width. Checkpoints bind the smearing method and width and reproduce the same occupations after resume. Periodic nonlocal forces also consume those occupations, so a converged smeared result yields the stationary free-energy force.

Both paths use reduced-coordinate Monkhorst-Pack meshes and 0.5|G + k|², including Bloch-phase local and nonlocal pseudopotential evaluation. run_periodic_band_structure reuses a converged SCF density and solves non-self-consistently along a high-symmetry path.

Stress, Relaxation, And Restart

The legacy finite-difference stress remains an orthorhombic teaching surface. The periodic stress oracle transports one converged variational state across full-rank cells, reevaluates every energy term on a fixed integer-G topology, and requires primary and doubled strain derivatives to agree. The periodic optimize_periodic_geometry workflow handles fixed-cell ions. The separate optimize_periodic_cell workflow composes cell-only or ion/cell steps with an enthalpy Armijo line search and an atomic accepted-cell checkpoint. Both outer workflows reuse accepted electronic state and reject failed trials.

The periodic workflow is current-verified for one displaced eight-atom Silicon crystal. It converged in seven accepted steps to a maximum force of 8.832e-5 Ha/bohr and a translation-aligned ideal-geometry error of 0.000312 A. This is a bounded fixed-cell result, not variable-cell or broad materials certification.

The same workflow is current-verified in one four-atom hexagonal 2H-Silicon cell. It converged in three accepted steps to a maximum force of 1.715e-5 Ha/bohr. Full-rank geometry is implemented across grids, reciprocal bases, Ewald, GTH operators, forces, fingerprints, and state metadata. Stress and variable-cell control pass deterministic oracles. A source-bound cell trajectory from 0.995 of the accepted scale converged after one accepted step at scale 0.9981142, with final pressure 2.66553e-6 Ha/bohr³. A fresh 35 Ha stress differed from the 25 Ha result by only 2.17019e-7 Ha/bohr³, closing the locked Pulay gate. This is a bounded 2H-Silicon result, not broad material certification.

Dense SCF restart files store density, orbitals, ion positions, cell lengths, spin metadata, and Γ k-point metadata for small-system continuation workflows.

Reference Validation

Reference comparison is intentionally static and lightweight. Fixtures are JSON summaries; QE/CP2K are not imported, built, or required in CI. The comparison helper records observed energy, expected energy, error, and pass/fail against a documented tolerance.

DFT/QM Platform Scope

get_dft_qm_scope_report() classifies local DFT/QM capability against CP2K and Quantum ESPRESSO reference families without changing the runtime dependency boundary.

Feature	Local Status	Reference Family
Plane-wave SCF core	verified for fixed-occupation bulk-Si EOS and one Fermi-Dirac fcc-Al EOS	CP2K Quickstep, QE PWscf
Full-rank fixed periodic cells	verified for one bounded 2H-Si relaxation and low-symmetry numerical oracles	CP2K cell matrix, QE CELL_PARAMETERS
UPF/GTH pseudopotentials and nonlocal projectors	GTH material evidence remains blocked; scalar norm-conserving UPF has source-bound execution and numerical gates but no material certification	QE UPF, CP2K GTH
Fixed-cell periodic geometry relaxation	verified for bounded orthorhombic and hexagonal Si workloads	CP2K MOTION/GEO_OPT, QE relax
Analytic periodic stress	verified for deterministic derivatives and one cutoff-converged 2H-Si material path	CP2K stress, QE stress
Variable-cell relaxation and restart	verified for one bounded 2H-Si cell-only path; deterministic coupled and resume gates pass	CP2K CELL_OPT, QE vc-relax
Static reference comparison	supported	static CP2K/QE fixture summaries
QM/MM force-environment orchestration	deferred	CP2K FORCE_EVAL/QMMM
PH/EPW/NEB/TDDFT/MPI/offload suite breadth	deferred	QE and CP2K production suites
Importing, wrapping, building, or running CP2K/QE	anti-goal	external executables

Plane-wave SCF has source-bound equation-of-state validation for bulk Silicon and simple-metal fcc Aluminum. Aluminum uses a matching GTH family, Fermi-Dirac convention, converged weighted k-point mesh, and Helmholtz free-energy definition. Broader chemistry stays proof-level, and the separate MLX-versus-QE PWscf cross-engine parity is still diagnostic, not closed.

dft_qm_scope_readiness_report() returns a shared readiness payload for these features. Deferred, anti-goal, and unknown features report blockers before any production-suite claim can be emitted.

The ordered path from this bounded capability to a general solid-state core is maintained in the DFT roadmap. That roadmap separates scientific protocol research from implementation and defines the material-level exit gate for every new capability.

Hot-Path Recommendation

Hamiltonian application remains the largest periodic DFT phase, but retained evidence does not justify another narrow custom Metal wrapper. One-dimensional and three-dimensional scatter/gather kernels improved isolated boundaries but did not clear the complete-run retention gate. A compiled local FFT wrapper also changed the convergence trajectory and regressed complete wall time.

The current route instead removes algorithmic work. Final-tolerance Davidson inverse FFTs are reused by density construction, while earlier adaptive cycles defer direct validation and perform only the inverse FFT needed for density. Further Hpsi work should likewise reduce FFT applications or useful vector-equivalents while preserving final direct-residual validation. Orthogonalization is the second measured target. The current measurements and rejected boundaries are maintained in the DFT performance decision ledger.