QHO σ_thermal table

QHO-backed sigma_thermal lookup for runtime ε_noise calibration.

atom_hifi.training.qho_sigma.load_qho_table()[source]

Load packaged QHO frequency table.

Default is the msq/kappa_stiff (stretch-mode, second-moment) estimator, not the centroid (full-DOS average) estimator — see ADR-0010: clean chemical-motif identity under centroid does not imply MLIP-training- adequate granularity, so the default now shares the stiffer moment used by the opt-in force-truth (microstate.reference: f_tol) path.

Return type:

dict

atom_hifi.training.qho_sigma.qho_table_version()[source]

Version tag of the packaged QHO table (_meta.version; 'unknown' if absent).

Saved in eps_noise_raw.npz and compared by the runner’s ε-cache check when sigma_source is qho_table.

Return type:

str

atom_hifi.training.qho_sigma.qho_sigma_from_omega_mev(element, omega_eff_meV, T)[source]

Return QHO RMS displacement σ (Å) for an element/frequency/temperature.

Parameters:
Return type:

float

atom_hifi.training.qho_sigma.qho_sigma_thermal(element, T=300.0)[source]

QHO-table RMS displacement σ_thermal (Å).

Parameters:
Return type:

float

atom_hifi.training.qho_sigma.resolve_sigma_thermal_targets(species, *, T, source='qho_table', config_override=None)[source]

Resolve per-element σ targets for thermal ε calibration.

Precedence (highest first): config_override (eps_noise.thermal.sigma_override), then the deprecated calibration.SIGMA_THERMAL_OVERRIDE (its warning is logged even when a config pin shadows it), then the source model. Pin values are exact Å for the current run temperature (not rescaled with T). Explicit pins skip the QHO-table lookup, so pinned species need no table entry. Every target must be finite and positive.

Parameters:
Return type:

dict[str, float]