DiffeRT integration¶
DiffeRT is a JAX-based differentiable ray tracer. It enumerates all paths up to a given bounce order by exhaustive search, which makes it exact but O(N^K) in scene triangles N and bounce order K. It works on any platform where JAX runs, including Windows.
Requirements¶
DiffeRT runs on CPU by default. JAX GPU acceleration is supported if a CUDA-compatible GPU and the jax[cuda] package are present, but it is not required.
High-level usage¶
paths_from_differt_scene() is the main entry point. It loads a Mitsuba/Sionna XML scene, runs ray tracing for each TX element and bounce order, and returns a PropagationPaths object ready for any AEGIS fidelity level.
import numpy as np
from aegis.integration.differt import paths_from_differt_scene
from aegis import DosimetryEngine
from aegis.tissue.dielectric import SKIN_28GHZ
paths = paths_from_differt_scene(
scene_path="scene.xml",
tx_positions=np.array([[0.0, 0.0, 10.0]]), # (M_ant, 3) [m]
rx_position=np.array([5.0, 0.0, 1.5]), # body position [m]
freq_hz=28e9,
max_bounces=3,
tx_power_dbm=30.0,
initial_polarisation="vertical",
)
engine = DosimetryEngine(SKIN_28GHZ)
result = engine.compute(body, paths, level=2)
The scene_path accepts the same Mitsuba XML format used by Sionna RT, so scenes created for Sionna work with DiffeRT without modification.
Low-level usage¶
If you already have DiffeRT path data, paths_from_differt() converts it directly to PropagationPaths without re-running the scene loading or path enumeration. This is useful when you want to control the DiffeRT scene setup yourself or integrate into an existing JAX pipeline.
from aegis.integration.differt import paths_from_differt
paths = paths_from_differt(
vertices=scene_vertices, # (N_triangles, 3)
normals=scene_normals, # (N_triangles, 3)
path_vertices=path_verts, # (N_paths, N_bounces+2, 3)
tx_positions=tx_pos, # (M_ant, 3)
freq_hz=28e9,
tx_power_dbm=30.0,
object_indices=obj_idx, # (N_paths, path_length)
material_indices=face_materials, # (N_triangles,)
material_n_tilde=n_tilde_list, # list of complex per material
initial_polarisation="vertical",
)
When object_indices and material_n_tilde are omitted, AEGIS assigns an arbitrary perpendicular polarisation to each path. This is sufficient for incoherent fidelity levels (0-6), where only \(|\boldsymbol{\psi}|^2\) enters the computation.
How psi is computed¶
For each path, AEGIS computes the E-field amplitude from free-space path loss:
where \(d\) is the total path length, \(Z_0 = 376.73\;\Omega\), and \(P_T\) is the transmit power in watts. The initial polarisation vector is set from the TX antenna orientation (vertical or horizontal).
At each reflection, the field is decomposed into TE and TM components. Fresnel reflection coefficients are applied using the surface normal and the material refractive index. The reflected field is reconstructed in the new propagation frame. Material properties default to concrete (\(\varepsilon_r = 5.31\), \(\sigma = 0.0326\) S/m at 28 GHz) when the scene provides no material assignments.
The arrival direction \(\hat{k}\) is taken from the last non-degenerate path segment. DiffeRT pads variable-length paths to the same array shape, so zero-length padding segments are skipped.
Element indexing¶
For multi-element TX arrays, each path is assigned to a TX element. If element_indices is not provided, AEGIS assigns each path to the nearest TX element by Euclidean distance from the path's first vertex. Pass explicit element_indices when paths are pre-sorted per element for best accuracy.
Platform compatibility¶
DiffeRT works on Linux, Windows, and macOS. It does not require a GPU. The JAX backend defaults to CPU. Set JAX_PLATFORM_NAME=gpu or call jax.config.update("jax_platform_name", "gpu") before importing to use GPU acceleration.
For city-scale scenes where the O(N^K) enumeration becomes too slow, switch to Sionna RT. See Sionna RT integration for details and a feature comparison table.
