Full three-dimensional P3 transport#
P3 approximates the angular radiance using all real spherical harmonics
through degree three. In three dimensions there are 16 coefficients: one of
degree zero, three of degree one, five of degree two, and seven of degree
three. This is distinct from SP3 transport model, whose two scalar equations simplify
the multidimensional angular structure. Neither approximation recovers the
full radiative-transfer solution in every configuration.
Angular representation and coefficients#
For real orthonormal harmonics \(Y_{lm}\) on the unit sphere,
MomentRT orders each degree by \(m=-l,\ldots,l\); negative orders use sine and positive orders cosine with the Condon–Shortley convention. The six coefficients of degrees zero and two form \(\mathbf c_e\); the ten coefficients of degrees one and three form \(\mathbf c_o\). Fluence is \(\Phi=\sqrt{4\pi}\,c_{00}\).
For each material’s normalized phase-function moments \(g_l\) and reduced scattering coefficient \(\mu_s'=\mu_s(1-g_1)\), the collision coefficient for degree \(l\) is
Thus \(\sigma_0=\mu_a\). For Henyey–Greenstein \(g_l=g^l\). Coefficients are evaluated per material at the source wavelength and sampled at tetrahedron centroids. No assumption that different materials share an anisotropy is needed. Internal refractive-index jumps are currently rejected.
Even-parity formulation#
Define the angular streaming matrices and diagonal collision matrices by
With no internal source, the odd moment equation can be eliminated locally:
Substitution into the even equations gives six coupled second-order equations, with the tensor coupling retained in three dimensions:
Each even moment uses the same continuous linear tetrahedral basis. The ten
odd coefficients are reconstructed at element centers; they can be
discontinuous across element or material boundaries. The reported
result.flux is the first angular moment reconstructed there. The six
nodal even coefficients are available in result.moments under names
such as l0_m0 and l2_m0. Their values follow the normalized harmonic
convention above; they are not themselves the SP3 composite moments.
Matched-index boundary and illumination#
For outward normal \(\mathbf n\), the current of the even test moments is \(\mathbf q_n=\sum_i n_i A_i^T\mathbf c_o\). Project the prescribed incoming radiance onto the incoming half sphere using the six weights \((\mathbf\Omega\cdot\mathbf n)Y_e\). For an isotropic incoming hemisphere with partial current \(I_{\rm in}\), its radiance is \(I_{\rm in}/\pi\). Define
Parity of the even and odd harmonics then gives the matched-index Marshak condition
The weak form adds \(B\) as a boundary mass matrix and \(2I_{\rm in}\mathbf b\) as the source load. Angular matrices are evaluated using Gauss–Legendre and equally spaced azimuthal quadrature; the chosen rules integrate the required P3 polynomial products. In planar geometry the azimuthally symmetric P3 coefficients reduce to the familiar two composite equations and Marshak coefficients documented in SP3 transport model.
All six faces currently require RobinBoundary(effective_reflectance=0).
Reflecting boundaries and index-mismatched or refracting interfaces are not
implemented for P3 and are rejected explicitly. Surface sources specify
incoming diffuse power already coupled into the sample; collimated fiber
propagation and entry losses are outside this model.
The outgoing face partial current is the reconstructed net normal current plus the incoming partial current. The existing contact detector integrates that escaping field over its aperture and applies its separate Lambertian angular-acceptance approximation. It does not yet integrate the reconstructed P3 radiance over a fiber’s acceptance cone, and does not model fiber modes or Fresnel transmission.
Numerical checks and limits#
Independent planar matrix-exponential slab solutions check the P3 reduction, boundary source, and spatial refinement. A three-dimensional nearly non-scattering cube is compared with an independent ballistic ray integral; its nonzero discrepancy is expected because P3 is a truncated angular model. Tests also exercise axis rotation, energy balance, spectral preparation, detector quadrature, and unsupported configurations. These cases do not establish accuracy for skin, all optical properties, or short source–detector separations. Compare the intended sensor geometry against Monte Carlo or a converged transport reference and assess spatial mesh convergence separately.
Higher-order angular truncations can produce negative reconstructed fluence or outgoing irradiance. MomentRT preserves these values and reports them; it does not silently clip them. P3 has six coupled nodal unknowns rather than SP3’s two and can require substantially more memory and computation. Actual cost depends on geometry, mesh, and sparse-factorization fill-in.
As one illustrative local measurement, a homogeneous 4 mm cube with a 9 × 9 × 9 mesh, uniform-face illumination, and direct sparse solves gave:
Model |
Solve time |
Process peak RSS |
|---|---|---|
P1 |
0.032 s |
161 MiB |
SP3 |
0.039 s |
164 MiB |
P3 |
0.662 s |
208 MiB |
Each model ran in a fresh process; peak RSS includes the interpreter and dependencies. These are single-run macOS measurements, not performance guarantees. Larger meshes can change both ratios and factorization memory substantially.
See Compare full P3 with SP3 and P1 for a runnable comparison. For the finite-element PN approach in optical tomography, see Mohan et al., Journal of Computational Physics 230 (2011), 7364–7383.