NYURay: Deterministic 3D Wireless Ray-Tracing Engine

• NYURay is a deterministic 3D ray-tracing engine that predicts site-specific wireless channel parameters across mmWave, sub-THz, and upper mid-band frequencies.
• It employs a hybrid shooting and bouncing rays method with image-based refinement to achieve directional multipath power prediction errors of less than 3 dB in various environments.
• NYURay supports advanced applications such as digital twin reconstruction, ISAC imaging, and beamforming evaluation, though its accuracy depends on detailed environmental maps and precise material calibration.

NYURay is a site-specific, deterministic 3D ray-tracing engine developed at NYU WIRELESS for wireless channel prediction across mmWave, sub-THz, and later upper mid-band and FR3 regimes. Its central purpose is to generate site-specific channel parameters such as powers, angles of arrival and departure, and time-of-flight of multipath components, so that channel characterization, deployment planning, localization, sensing, and wireless digital-twin studies can be conducted without relying exclusively on exhaustive measurement campaigns. Across its calibration and extension papers, NYURay is presented both as a calibrated propagation simulator and as a reusable physics engine for tasks including location calibration, automated digital-twin reconstruction, ISAC imaging, and sub-THz beamforming evaluation (Kanhere et al., 2023, Kanhere et al., 2024, Ying et al., 29 Jul 2025, Ying et al., 13 Feb 2026).

1. Scope, environments, and research context

NYURay was introduced as a deterministic 3D ray tracer specifically tailored to mmWave and sub-THz frequencies, explicitly calibrated at 28 GHz, 73 GHz, and 140/142 GHz in indoor office, outdoor urban microcell, and factory environments. The indoor environments include open-plan office floors with cubicles, drywall, glass, offices, and classrooms; the outdoor environments include Manhattan and Brooklyn campus urban spaces; and the industrial environments include multi-company office/factory space, electronics manufacturing, warehouse space, and prototyping space. Its stated purposes are channel characterization, system performance evaluation for 5G/6G communications and sensing, and deployment planning through site-specific prediction of channels, rather than purely stochastic modeling (Kanhere et al., 2023).

Subsequent work broadened this scope. At upper mid-band frequencies, NYURay was used at 6.75 GHz and 16.95 GHz in a dense urban microcell around NYU Tandon, with outputs aligned to 3GPP TR 38.901 channel metrics and point-data formats, and with explicit use for wireless digital twins, coverage analysis, and validation against measured omnidirectional PDPs, path loss, delay spread, and angular spreads (Ying et al., 29 Jul 2025). In a factory setting, HoRAMA used NYURay as the downstream propagation engine for 6.75 GHz and 16.95 GHz dual-band simulations over a 700 square meter factory, with the reconstruction pipeline designed specifically to satisfy NYURay’s geometric and material requirements (Ying et al., 13 Feb 2026).

This broader trajectory suggests that NYURay functions less as a single fixed-frequency simulator than as a family of site-specific deterministic workflows. A plausible implication is that its identity is defined primarily by calibrated geometry-aware propagation prediction, rather than by any one material model or interaction subset.

2. Propagation model and computational structure

In its original mmWave and sub-THz calibration formulation, NYURay models a received multipath component $j$ with a power expression in dB of the form

$$\begin{aligned} P_{j,R} [\text{dBm}] &= P_{j,TX} [\text{dBm}] + G_{j,T} [\text{dBi}] + G_{j,R} [\text{dBi}] \ &\quad - FSPL(d_j, f) [\text{dB}] \ &\quad - \sum_{i=1}^{N}\left(w^{pen}_{i,j}L^{pen}_{i} + w^{ref}_{i,j}L^{ref}_{i}\right) [\text{dB}], \end{aligned}$$

where the attenuation beyond free-space loss is a linear combination of material-specific reflection and penetration losses, weighted by the number of reflections and penetrations along the path. In that formulation, dominant mechanisms at mmWave and sub-THz are specular reflection, penetration, and diffuse scattering, while diffraction is treated as negligible and ignored (Kanhere et al., 2023).

The 2024 calibration paper gives a more explicit algorithmic description. NYURay uses a hybrid Shooting and Bouncing Rays front-end combined with an image-based refinement step. Rays are launched from the transmitter using a tessellated icosahedron; a reception sphere is used to detect candidate rays that reach the receiver; and image-theoretic reconstruction is then used to recover exact path geometry. Reflections are allowed up to 5th order, penetration follows a Wall Attenuation Factor model, and diffuse scattering uses the directive scattering dual-lobe model of Degli-Esposti with fixed parameters

$$\Lambda = 0.8,\quad \alpha_i = \alpha_R = 10,\quad S = 0.1.$$

Calibration experiments reported runtimes of about 93 ms per Tx–Rx location without scattering and about 1.1 s per location with scattering on an Intel i7-3770 with 16 GB RAM (Kanhere et al., 2024).

Later upper mid-band and FR3 work adopts a field-based formulation. For each path $p$, the electric field at the receiver is written as

$$\mathbf{E}_p(\mathbf{r}_R) = \frac{\lambda}{4\pi}\,\frac{e^{-j k d_p}}{d_p}\,\mathbf{F}_R \,\mathbf{T}_p \,\mathbf{F}_T \,\mathbf{E}_0,$$

with total field obtained by coherent summation over LOS, reflected, transmitted, diffracted, and scattered paths. In the 6.75/16.95 GHz urban validation study, reflection and diffraction were enabled, while penetration and scattering were disabled for that outdoor UMi scenario; in the HoRAMA factory study, reflection, penetration, and diffraction were enabled, with material-dependent Fresnel coefficients computed from ITU-R P.2040 electromagnetic parameters (Ying et al., 29 Jul 2025, Ying et al., 13 Feb 2026).

A common misconception is that NYURay has a single immutable propagation model. The record in fact shows a stable deterministic ray-tracing framework whose enabled mechanisms and material parameterizations vary by band, environment, and validation objective.

3. Calibration methodology and material parameterization

The methodological signature of the original NYURay papers is the assumption of angle-independent reflection loss per material. Instead of Fresnel coefficients that depend on incidence angle, polarization, and complex permittivity, NYURay assigns each material type $i$ a scalar reflection loss $L^{ref}_i$ and penetration loss $L^{pen}_i$, both in dB and independent of angle in the calibration equations. With

$$A_j = P_{j,TX} + G_{j,T} + G_{j,R} - FSPL(d_j,f) - P_{j,meas},$$

$$\mathbf{w}_j = [w^{pen}_{1,j}, \dotsc, w^{pen}_{N,j}, w^{ref}_{1,j}, \dotsc, w^{ref}_{N,j}],$$

and

$$\begin{aligned} P_{j,R} [\text{dBm}] &= P_{j,TX} [\text{dBm}] + G_{j,T} [\text{dBi}] + G_{j,R} [\text{dBi}] \ &\quad - FSPL(d_j, f) [\text{dB}] \ &\quad - \sum_{i=1}^{N}\left(w^{pen}_{i,j}L^{pen}_{i} + w^{ref}_{i,j}L^{ref}_{i}\right) [\text{dB}], \end{aligned}$$0

the calibration objective becomes

$$\begin{aligned} P_{j,R} [\text{dBm}] &= P_{j,TX} [\text{dBm}] + G_{j,T} [\text{dBi}] + G_{j,R} [\text{dBi}] \ &\quad - FSPL(d_j, f) [\text{dB}] \ &\quad - \sum_{i=1}^{N}\left(w^{pen}_{i,j}L^{pen}_{i} + w^{ref}_{i,j}L^{ref}_{i}\right) [\text{dB}], \end{aligned}$$1

which yields the closed-form least-squares solution

$$\begin{aligned} P_{j,R} [\text{dBm}] &= P_{j,TX} [\text{dBm}] + G_{j,T} [\text{dBi}] + G_{j,R} [\text{dBi}] \ &\quad - FSPL(d_j, f) [\text{dB}] \ &\quad - \sum_{i=1}^{N}\left(w^{pen}_{i,j}L^{pen}_{i} + w^{ref}_{i,j}L^{ref}_{i}\right) [\text{dB}], \end{aligned}$$2

This converts material calibration into a linear least-squares problem over many measured multipath components and is presented as the reason the method is both low-complexity and analytically tractable (Kanhere et al., 2023, Kanhere et al., 2024).

The calibrated material values are explicitly frequency- and environment-specific. Representative examples include drywall in an indoor office with reflection loss 6.1 dB and penetration loss 4.0 dB at 28 GHz, and reflection loss 9.9 dB and penetration loss 9.2 dB at 140/142 GHz; standard glass in the same indoor setting with 3.5 dB reflection and 3.2 dB penetration at 28 GHz, and 24.5 dB reflection and 7.2 dB penetration at 140/142 GHz; and granite outdoors with reflection loss 6.9 dB at 28 GHz, 5.6 dB at 73 GHz, and 13.1 dB at 140/142 GHz. The reported overall trend is that reflection and penetration losses generally increase with frequency, with the specific note that foliage loss at 140 GHz was lower than at 73 GHz in the outdoor campaigns, likely due to seasonal differences (Kanhere et al., 2024).

Upper mid-band work adds a distinct calibration layer: site-specific location calibration. Because smartphone GPS introduced 5–10 m absolute position errors, the 6.75/16.95 GHz study optimized transmitter and receiver locations by matching simulated and measured omnidirectional PDPs. The loss combined peak matching, unmatched-peak penalties, shape mismatch, and a distance regularizer, and the optimization used alternating minimization with coarse grid search, fine grid refinement, and Powell’s method. The reported improvement in TX-RX location accuracy was 42.3% for line-of-sight and 13.5% for non-line-of-sight scenarios (Ying et al., 29 Jul 2025).

4. Measurement campaigns, datasets, and validation results

The principal mmWave and sub-THz calibration corpus spans nine measurement campaigns between 2012 and 2022. These comprise indoor office measurements at 28 GHz and 140/142 GHz, outdoor urban microcell measurements at 28 GHz, 73 GHz, and 140/142 GHz, and four distinct factory campaigns at 140/142 GHz. Measurements used sliding-correlation, super-heterodyne channel sounders with 800 MHz RF bandwidth at 28 GHz and 1 GHz RF bandwidth at 73 and 140/142 GHz, together with high-gain directional horn antennas mounted on steerable gimbals for azimuth and elevation scanning (Kanhere et al., 2023, Kanhere et al., 2024).

After calibration, NYURay’s primary validation target was individual directional multipath power. In the 2023 and 2024 calibration studies, the standard deviation in the error of the directional multipath power predicted by the ray tracer compared to the directional measured power was less than 3 dB in indoor office environments and less than 2 dB in outdoor and factory environments. The detailed absolute-error standard deviations were 2.7 dB and 2.8 dB for indoor 28 GHz and 140/142 GHz, 2.0 dB, 1.6 dB, and 1.7 dB for outdoor 28 GHz, 73 GHz, and 140/142 GHz, and 2.0 dB, 1.9 dB, 1.7 dB, and 2.3 dB across factories A–D at 140/142 GHz (Kanhere et al., 2023, Kanhere et al., 2024).

Secondary metrics were not directly optimized and were therefore more revealing of model incompleteness. The 2024 calibration paper reports that NYURay tends to underpredict mean RMS delay spread by about 20% and angular spreads by about 10% on average, attributing this primarily to incomplete environmental maps and the absence of many weak, late scatterers in the CAD models. The 2023 paper similarly notes that some weaker later multipath components are not predicted, especially those arising from complex scattering or small objects not included in the environmental model (Kanhere et al., 2023, Kanhere et al., 2024).

At upper mid-band frequencies, validation was performed across 18 TX-RX locations in a dense urban microcell. Path-loss exponent deviations between measurements and NYURay were under 0.14 for both LOS and NLOS at 6.75 GHz and 16.95 GHz. The same study states that NYURay underestimates delay spread and angular spreads, but that their cumulative distributions remain statistically similar after outlier filtering, with KS-test agreement improving substantially once the most pathological RT/measurement mismatches are removed (Ying et al., 29 Jul 2025).

5. Extensions beyond calibration

NYURay has increasingly been used as a platform for site-specific wireless digital-twin workflows. HoRAMA, introduced as “Holistic Reconstruction with Automated Material Assignment,” is expressly designed to produce NYURay-compatible 3D models from RGB video. It integrates MASt3R-SLAM dense point cloud generation with vision LLM-assisted material assignment, then exports watertight, manifold, material-labeled Mitsuba XML meshes for NYURay. In a 700 square meter factory, NYURay predictions using HoRAMA-generated models achieved a 2.28 dB RMSE for matched multipath component power predictions, compared with 2.18 dB for a manually created 3D model baseline, while reducing 3D reconstruction time from two months to 16 hours (Ying et al., 13 Feb 2026).

NYURay has also been used as an idealized CSI generator for sensing and imaging. The ISAC imaging paper at 6.75 GHz treats per-path NYURay outputs

$$\begin{aligned} P_{j,R} [\text{dBm}] &= P_{j,TX} [\text{dBm}] + G_{j,T} [\text{dBi}] + G_{j,R} [\text{dBi}] \ &\quad - FSPL(d_j, f) [\text{dB}] \ &\quad - \sum_{i=1}^{N}\left(w^{pen}_{i,j}L^{pen}_{i} + w^{ref}_{i,j}L^{ref}_{i}\right) [\text{dB}], \end{aligned}$$3

as ground-truth per-path CSI parameters, and maps each resolvable multipath component into an equivalent reflection point through a constrained two-segment optimization. The paper states that it provides the first demonstration of multi bounce ISAC imaging using wireless ray tracing at 6.75 GHz, and qualitatively reconstructs trees, metallic cubes and plates, and a Tesla vehicle as dense three dimensional point clouds (Bazzi et al., 8 Sep 2025).

At 108 GHz, NYURay serves as the site-specific channel generator for evaluating hybrid beamforming with liquid crystal antennas and liquid neural networks in an urban scenario covering Brooklyn MetroTech Commons. Channels are generated with $$\begin{aligned} P_{j,R} [\text{dBm}] &= P_{j,TX} [\text{dBm}] + G_{j,T} [\text{dBi}] + G_{j,R} [\text{dBi}] \ &\quad - FSPL(d_j, f) [\text{dB}] \ &\quad - \sum_{i=1}^{N}\left(w^{pen}_{i,j}L^{pen}_{i} + w^{ref}_{i,j}L^{ref}_{i}\right) [\text{dB}], \end{aligned}$$4 rays, up to 5 reflections, and propagation modes reflection, penetration, and diffraction. On those NYURay-generated channels, the proposed method achieves an 88.6% spectral efficiency gain over the learning-aided gradient descent baseline and 1.9 times higher SE than the 3GPP TR 38.901 standard antenna model (Wang et al., 8 Apr 2026).

These later works shift NYURay from a calibration target to an infrastructural component. This suggests that the simulator’s scientific importance lies not only in reproducing measured channels, but also in supplying geometry-consistent synthetic data for inverse problems, algorithm design, and digital-twin maintenance.

6. Limitations, methodological distinctions, and interpretation

The most persistent limitation across NYURay studies is environmental incompleteness. The 28/73/140/142 GHz calibration papers explicitly attribute missing weak, late multipath components and underpredicted delay and angular spreads to rudimentary maps lacking small objects and clutter detail. The upper mid-band urban validation reaches a similar conclusion, noting that diffuse scattering, façade micro-geometry, dynamic scatterers, and exact environmental detail are not fully represented in the simulated scene (Kanhere et al., 2023, Ying et al., 29 Jul 2025).

A second limitation is that different NYURay studies make different physics choices. The original calibration work assumes angle-independent reflection loss and penetration loss per material and neglects diffraction because it is treated as negligible at mmWave and sub-THz in the considered environments. By contrast, the upper mid-band urban validation and the HoRAMA factory study use Fresnel-based reflection and transmission coefficients derived from ITU-R P.2040 material parameters and include diffraction. This is not a contradiction so much as a methodological distinction between a low-complexity calibration framework and later higher-fidelity digital-twin workflows (Kanhere et al., 2024, Ying et al., 13 Feb 2026).

A third limitation concerns idealization in downstream applications. The ISAC imaging framework assumes perfect per-path knowledge, no hardware impairments, infinite resolution, and exact environment geometry, because NYURay directly provides the “true” angles, delays, and complex gains. The 108 GHz beamforming study likewise uses NYURay-generated channel snapshots and then imposes channel estimation errors algorithmically, rather than deriving Doppler and temporal evolution directly from NYURay itself. These assumptions are appropriate for simulation studies, but they delimit how directly one can transfer the reported sensing or beamforming performance to deployment (Bazzi et al., 8 Sep 2025, Wang et al., 8 Apr 2026).

An objective synthesis of the literature is therefore twofold. First, NYURay is a rigorously validated site-specific ray tracer whose strongest demonstrated capability is directional multipath-power prediction with low error across indoor, outdoor, and factory measurements. Second, its predictive fidelity is conditional on map quality, material assignment, and the particular propagation model chosen for the study at hand. Within those conditions, the published record presents NYURay as a calibrated propagation core for channel modeling, wireless digital twins, sensing, localization, and sub-THz system evaluation.