Wannier90 compatibility

This page is the contract: which binaries and keywords are covered, which conventions are followed, and every known behavioural difference from the reference implementation.

Binaries

Keyword policy

The .win parser is strict: unknown keywords are an error with a did-you-mean suggestion, checked against a catalogue generated from the reference parser's own source. Keywords fall into three classes:

• Supported — consumed with reference semantics (the large majority: the wannierise, disentanglement, plotting, transport and all postw90 module keywords).
• Ignored — cosmetic/reporting controls (num_print_cycles, iprint, timing_level, …): accepted silently.
• Recognised but unsupported — valid wannier90 keywords whose feature is not implemented here; these warn once and are ignored, so reference input decks still run.

Notable keyword semantics matched to the reference (discovered from its source, not its documentation):

• berry_task/gyrotropic_task match by substring (eval_shc enables shc; glued -C-dos-D0 works); task lists split on +.
• The global interpolation mesh keyword is kmesh; module-specific *_kmesh override it.
• kubo_eigval_max defaults to dis_froz_max + 2/3 when a frozen window exists, else max(ε) + 2/3.
• BoltzWann's TDF energy window is the disentanglement window ± 0.2 eV; its DOS range defaults to the eigenvalue range ± 0.6667 eV.
• dis_proj_min/dis_proj_max (PDWF): proj_min is the pool threshold (discard below), proj_max the freeze threshold.
• Projection blocks accept both ; and , as orbital separators (O:s,p ≡ O:s;p).

Conventions

• Bohr radius: CODATA2006 (0.52917720859 Å) — the reference default. With it, validated spreads match reference prints to ≤ 5·10⁻¹⁰.
• use_ws_distance defaults to true (as in postw90) — an ~11% effect on transport coefficients in the validation set; disable explicitly to compare with older runs.
• transl_inv_full phases and corrections act on the expanded minimal-image R-set (the reference's operator_wigner_setup), matching its numerics exactly.
• Γ-only mode runs the real-orthogonal reference algorithm (wann_main_gamma) and returns exactly real gauges; the model-level API can optionally minimise over the full unitary group instead (marginally lower Ω, complex gauge).

Known differences and caveats

Everything on this list is deliberate and documented; nothing else is known to differ.

• tran_lcr (lead–conductor–lead transport) is not implemented — the one reference feature that is out of scope (no shipped oracle; dedicated NEGF codes serve this use case). transport_mode = bulk is fully supported and validated.
• Upstream bug (reported herein): tran_read_ht — the reference v4-dev binary segfaults on transport-from-file inputs; our read_ht/transport_bulk path works and was validated against the reference's full pipeline instead.
• Upstream bug: Ryoo SHC with shc_gamma = 2, 3 — the reference accumulates the spin σ-rotation across polarisations in get_SAA_R/get_SBB_R (getoper.F90), mixing spin components. We compute clean per-component values; results agree for `shcgamma = 1`.
• Stengel–Spaldin functional: the SS objective surface is a near-flat valley, so the stopping point depends on the convergence criterion even between two wannier90.x runs (13.845 vs 13.312 Ų on the reference test with different conv_tol). We validate state functions (objective and gradient at shared gauges, to 9–10 digits) rather than the stopping point, and run the reference optimiser so trajectories track.
• Injection current has no Fortran reference; it is cross-validated against WannierBerri on a shared tight-binding model. Agreement is at the ~10⁻⁴ level, limited by the two codes' degenerate-state regularisation conventions.
• BoltzWann TDF shows ≤ 10⁻⁴-relative deviations at isolated band-crossing energies (degenerate-band velocity conventions); all other BoltzWann outputs are identical to the last printed digit.
• kdotp orders 1–2 are gauge-dependent (the reference's own test checks only order 0, which matches byte-identically); magnitudes agree to ~10⁻⁶ relative.
• .wout/stdout text: the CLI writes a faithful .wout for the wannierise stage; postw90.jl prints a compact log rather than replicating the .wpout prose (its .dat files are the validated deliverables).

Feature-by-feature validation anchors

The complete list of what was validated against which oracle, with numbers, lives in Validation; implementation-depth notes (conventions, oracle anchors, exact Fortran formats) are in the repository's docs/reference-notes/ directory — one note per algorithm, including the parity audit (parity-audit-2026.md) and the gap-closing slate (remaining-gaps-2026-07.md).