Struct PrecomputedTransmissionModel 

pub struct PrecomputedTransmissionModel {
    pub cross_sections: Arc<Vec<Vec<f64>>>,
    pub density_indices: Arc<Vec<usize>>,
    pub energies: Option<Arc<Vec<f64>>>,
    pub instrument: Option<Arc<InstrumentParams>>,
    pub resolution_plan: Option<Arc<ResolutionPlan>>,
    pub sparse_cubature_plan: Option<Arc<SparseEmpiricalCubaturePlan>>,
    pub sparse_scalar_plan: Option<Arc<ScalarSurrogatePlan>>,
    pub work_layout: Option<Arc<WorkingGridLayout>>,
}

Transmission model backed by precomputed Doppler-broadened cross-sections.

The expensive physics steps (resonance → σ(E), Doppler broadening) are computed once and stored. Each evaluate() call performs Beer-Lambert and, when instrument is present, resolution broadening on the total transmission:

T(E) = R ⊗ exp(−Σᵢ nᵢ · σ_{D,i}(E))

Issue #442: resolution broadening is applied to T(E) after Beer-Lambert, not to σ(E) before.

Construct via nereids_physics::transmission::broadened_cross_sections, then wrap in Arc so the same precomputed data is shared read-only across all rayon worker threads.

Fields

cross_sections: Arc<Vec<Vec<f64>>>

Doppler-broadened cross-sections σ_D(E) per isotope, shape [n_isotopes][n_grid_energies].

The grid these σ live on is determined by work_layout:

• work_layout is Some (Gaussian resolution → auxiliary extended grid): σ live on the working grid, i.e. work_layout.energies, with n_grid_energies == work_layout.energies.len(). evaluate() / analytical_jacobian() apply Beer-Lambert + resolution on this working grid and extract the data points LAST via work_layout.extract(..) — matching forward_model (issue #608).
• work_layout is None (tabulated resolution, or no resolution): the working grid IS the data grid, so σ live on the data grid (energies), with n_grid_energies == energies.len(). No extraction is needed and the surrogate fast paths + data-grid resolution_plan behave exactly as before.

density_indices: Arc<Vec<usize>>

Mapping: params[density_indices[i]] is the density of isotope i.

Wrapped in Arc so that parallel pixel loops can share one copy via cheap reference-count increments instead of deep-cloning per pixel.

Kept pub (not pub(crate)) because the Python bindings (nereids-python) construct and access this field directly.

energies: Option<Arc<Vec<f64>>>

Energy grid (eV), required for resolution broadening. None when resolution is disabled — Beer-Lambert only.

instrument: Option<Arc<InstrumentParams>>

Instrument resolution parameters. When Some, resolution broadening is applied to the total transmission after Beer-Lambert in evaluate().

resolution_plan: Option<Arc<ResolutionPlan>>

Optional pre-built broadening plan for (energies, resolution).

When a caller builds the plan once (e.g. spatial dispatch for a grid shared across every pixel) and passes it via with_resolution_plan, evaluate() and analytical_jacobian() skip the per-call kernel-interp / bracket / trap-weight work and reduce each broadening call to a gather + multiply-add. None ⇒ fall back to the per-call broadening path, byte- identical output.

sparse_cubature_plan: Option<Arc<SparseEmpiricalCubaturePlan>>

Optional sparse empirical cubature plan (epic #472).

When the plan is present AND its target_energies match this model’s energy grid AND cubature.k() == n_density_params AND no temperature / energy-scale fitting is active, the evaluate() / analytical_jacobian() fast path calls cubature.forward_and_jacobian(n) directly instead of exp(-Σ n σ) + apply_resolution. Any guard failure falls back to the exact path, so the default behaviour is byte-identical to main.

sparse_scalar_plan: Option<Arc<ScalarSurrogatePlan>>

Optional scalar (k = 1) surrogate plan (epic #472, PR #475).

Mutually exclusive with sparse_cubature_plan in practice — the cubature dispatch fires only for k ≥ 2 and the scalar plan only for k == 1. The type alias ScalarSurrogatePlan = ScalarChebyshevPlan is kept as a stable public name so a future scalar surrogate can swap in without touching this field or any dispatch call site. PR #475 picked Chebyshev-in-density over Lanczos Gauss quadrature after a real-VENUS bench-off (Chebyshev won on both the accuracy and wall-time axes; see nereids_physics::surrogate module docs).

work_layout: Option<Arc<WorkingGridLayout>>

Working-grid layout matching cross_sections.

Issue #608: when cross_sections is stored on the auxiliary extended grid (Gaussian resolution), this maps the working grid back to the data grid so evaluate() / analytical_jacobian() apply resolution on the working grid and extract the data points last. None ⇒ the working grid is the data grid (tabulated / no resolution): Beer-Lambert and resolution run directly on energies and no extraction is needed, which keeps the surrogate fast paths and the data-grid resolution_plan byte-identical to before.

Trait Implementations

impl FitModel for PrecomputedTransmissionModel

fn analytical_jacobian( &self, params: &[f64], free_param_indices: &[usize], y_current: &[f64], ) -> Option<FlatMatrix>

Analytical Jacobian for the Beer-Lambert transmission model.

Without resolution: T(E) = exp(-Σᵢ nᵢ · σᵢ(E)) ∂T/∂nᵢ = -σᵢ(E) · T(E)

With resolution (R is a linear operator): T_obs(E) = R[T](E) = R[exp(-Σᵢ nᵢ · σᵢ)](E) ∂T_obs/∂nᵢ = R[-σᵢ(E) · T(E)]

For grouped isotopes sharing density parameter N_g: ∂T_obs/∂N_g = R[-(Σ_{i∈g} σᵢ(E)) · T(E)]

fn evaluate(&self, params: &[f64]) -> Result<Vec<f64>, FittingError>

Evaluate the model for the given parameters.

impl ForwardModel for PrecomputedTransmissionModel

fn predict(&self, params: &[f64]) -> Result<Vec<f64>, FittingError>

Predict model output for the given parameter vector.

fn jacobian( &self, params: &[f64], free_param_indices: &[usize], y_current: &[f64], ) -> Option<Vec<Vec<f64>>>

Analytical Jacobian (column-major layout).

fn n_data(&self) -> usize

Number of data points in the model output.

fn n_params(&self) -> usize

Number of parameters (total, including fixed).