Acoustic Radiance Transfer Overview

• Acoustic Radiance Transfer (ART) is a framework that models sound propagation via directional energy transfer between discretized surface patches, emphasizing material properties and geometry.
• ART applies radiative transfer theory to acoustics by integrating techniques like delay networks, modal approximations, and neural operators for efficient simulation of late reverberation and audio-visual synthesis.
• ART bridges geometric acoustics and wave-based corrections, offering efficient transport-based approximations while acknowledging limitations in low-frequency accuracy and boundary-layer phenomena.

Acoustic Radiance Transfer (ART) is a family of techniques that recast sound propagation in terms of radiance-like transfer functions, closely analogous to radiance transfer in graphics. In room acoustics, ART solves the time-dependent acoustic rendering equation through discretization, modeling the time- and direction-dependent energy exchange between surface patches given with flexible material properties; more broadly, it denotes an energy-transport viewpoint in which directional acoustic energy is propagated, scattered, delayed, and observed through transfer operators rather than through direct solution of the full wave equation at runtime (Lee et al., 19 Sep 2025, Tang et al., 2020). ART therefore occupies a middle ground between geometric acoustics, radiative transfer theory, precomputed transfer, and more recent neural or differentiable surrogates, with applications ranging from late reverberation and interactive auralization to cross-modal audio-visual scene synthesis (Scerbo et al., 2023, Brunetto et al., 2024).

1. Conceptual and mathematical basis

ART is rooted in geometrical acoustics. The basic assumptions are that sound is modeled as rays rather than waves, that room geometry is large compared to the acoustic wavelength, and that the transported quantity is acoustic radiance, i.e. energy flux per unit projected area per unit solid angle. In the room-acoustics formulation, the underlying continuous model is the Room Acoustics Rendering Equation (RARE), a Fredholm integral equation of the second kind that is structurally identical to the rendering equation in radiometry and optics (Scerbo et al., 2023, Lee et al., 19 Sep 2025).

Radiance is defined as

$$L(x,v,t)\; [\mathrm{W}/(\mathrm{m}^2\,\mathrm{sr})]$$

for a point $x$ on the boundary surface, in direction $v$, at time $t$. For two surface elements $A_1$ and $A_2$, the transferred power can be written as

$$P(A_1, A_2, t) = \iint_{A_2} \iint_{A_1} L\big(x, v(x|x'), t\big)\; d\omega_x(x')\; dA_{\text{p}, x'}(x).$$

The continuous RARE then expresses outgoing radiance at one surface point as the sum of source-originated radiance and reflected contributions from all other surface points: $$L\big(x', v(x'|x''), t\big) = \iint_A L\big(x, v(x|x'), t\big)\; R\big(x, x', v(x'|x'')\big)\; dA(x) + L_0\big(x', v(x'|x''), t\big).$$

The reflection kernel is factored into material, geometry, visibility, and propagation terms: $$R\big(x, x', v(x'|x'')\big) = \rho\big(x' \mid v(x'|x), v(x'|x'')\big)\; G(x, x')\; V(x, x')\; D(x,x').$$ Here $\rho$ is the BRDF or, in the more general DART formulation, a BSDF that allows reflection and transmission; $x$0 is the geometric factor; $x$1 is visibility; and $x$2 is the propagation operator carrying time delay and air absorption. DART states the same structure in explicitly time-dependent form, with a kernel

$x$3

where

$x$4

This formulation makes the governing abstractions explicit. ART does not propagate acoustic pressure phase as the primary state; it propagates acoustic radiance or energy density. That distinction is central. In this sense, ART is closer to a transport theory than to a wave solver, even when later work derives ART-relevant transport equations from first principles.

2. Discretization, transfer operators, and system representations

Classical ART discretizes boundary surfaces into patches and time into samples. If the boundary is partitioned as

$x$5

then patch-to-patch radiance is defined as

$x$6

The reflection kernel is likewise averaged over patch triples $x$7: $x$8 The resulting discrete RARE is

$x$9

When propagation over a patch pair is approximated as a single delay plus a gain, this becomes

$v$0

where $v$1 is a discrete delay and $v$2 is the ART scattering matrix. Physical BRDFs imply an energy constraint,

$v$3

Source injection and receiver detection are explicit parts of the model: $v$4

$v$5

These equations define the standard ART pipeline: a physically based energy-transfer model between patches, driven by sources and observed at receivers (Scerbo et al., 2023).

More recent work generalizes this discretization in two directions. DART discretizes not only patches but also outgoing directions, producing a patch-direction radiance state $v$6 and a discrete kernel $v$7. Its discrete acoustic rendering equation is

$v$8

MoD-ART, by contrast, starts from Time-Domain ART as a recursive discrete-time LTI system whose impulse response is the echogram, i.e. the squared room impulse response. In that representation, ART has a state vector of energy signals traversing volumetric paths, a feedback matrix encoding reflection and scattering between paths, a diagonal delay matrix encoding propagation and air absorption, and source/listener gain matrices encoding source-to-path and path-to-listener coupling (Scerbo et al., 2024, Lee et al., 19 Sep 2025).

A recurrent misconception is that ART is simply “ray tracing with delays.” The discrete equations show otherwise. ART is a structured transfer system over a graph of patch or path states, with explicit energy recursion, source injection, and receiver detection. Ray tracing may be used to estimate visibility, injection, or form factors, but the governing representation is radiance transfer.

3. Radiative-transfer derivations and wave-based corrections

Several works connect ART-like transport directly to the acoustic wave equation. In the weak-scattering regime, the Wigner transform of the wave solution yields a radiative transfer equation for directional energy density,

$v$9

with scattering coefficients

$t$0

$t$1

In ART notation, $t$2 is the radiance or directional energy density, the transport term is ray-based propagation, and the integral term is the scattering operator (Messaoudi et al., 2022, Messaoudi et al., 2023).

The half-space and slab analyses show that standard reflective boundary conditions capture only part of the boundary physics. In both settings, the bulk radiative transfer equation is accompanied by a boundary-layer correction supported within one wavelength of the boundary. For homogeneous Neumann boundary conditions, the boundary effect yields a doubling of the intensity; for homogeneous Dirichlet boundary conditions, it yields a canceling of that intensity (Messaoudi et al., 2022). In slabs, a second effect appears inside the domain: constructive interference of waves that have reflected at least twice on the boundaries only impacts the coherent part of the energy density and produces weak localization along interior planes (Messaoudi et al., 2023). This suggests that purely geometric ART can miss boundary-layer interference and weak-localization phenomena even when its bulk transport model is appropriate.

A second correction concerns the underlying wave equation itself. “Radiative transfer of acoustic waves in continuous complex media: Beyond the Helmholtz equation” derives transport from an acoustic wave equation with fluctuations of both density and compressibility. The random potential then contains both a scalar term and an operator term. Discarding the operator term has a significant impact on the phase function $t$3, anisotropy factor $t$4, and especially the transport mean-free path $t$5: when the scalar and operator terms have identical amplitudes, the discrepancy on the transport mean-free path is around $t$6 in the low-frequency regime, and still above $t$7 for $t$8 no matter how weak fluctuations of the disorder are (Baydoun et al., 2016). The immediate ART implication is that a scalar Helmholtz-derived transport kernel can substantially bias radiance transfer in heterogeneous media.

These wave-derived results do not replace classical room-acoustic ART. Rather, they delimit its validity. ART is exact neither as a wave model nor as a boundary-interference model; it is a transport approximation whose coefficients, boundary laws, and corrections may themselves require wave-based derivation.

4. Hybrid, neural, and differentiable ART variants

Recent work extends ART from explicit transfer matrices toward learned, hybrid, or differentiable operators. These systems are not uniform in method, but they preserve the ART premise that acoustic behavior can be represented as a transfer field conditioned on geometry, materials, directions, and source-listener configuration (Brunetto et al., 2024, Tang et al., 2020, Jin et al., 6 Jun 2025, Lee et al., 19 Sep 2025).

NeRAF formulates an acoustic field jointly with a radiance field. Its neural acoustic field is

$t$9

and, under the SoundSpaces simplification,

$A_1$0

The system conditions RIR prediction on a 7-channel voxel grid extracted from a Nerfacto radiance field, where channels consist of view-independent color, alpha or density, and voxel-center coordinates. A ResNet3D produces a 2048-dimensional scene feature, and an MLP predicts log-magnitude STFT coefficients of binaural RIRs. Conceptually, this is an implicit, data-driven version of radiance or ray transfer for sound: rather than tracing sound rays or solving the wave equation, it learns how sound energy is distributed in space from examples of room impulse responses, conditioned on the same 3D scene that a NeRF uses for light (Brunetto et al., 2024).

“Learning Acoustic Scattering Fields for Dynamic Interactive Sound Propagation” occupies a complementary position. It uses ray tracing for free-space propagation and large-boundary reflections, while learning a localized acoustic scattering field around arbitrary 3D objects. The scattered field on a spherical offset surface is encoded with spherical harmonics,

$A_1$1

with $A_1$2, giving $A_1$3 coefficients. PointNet maps a point cloud of $A_1$4 surface samples to those coefficients. Average normalized reproduction error over 10k test objects is $A_1$5 at 125 Hz, $A_1$6 at 250 Hz, $A_1$7 at 500 Hz, $A_1$8 at 1000 Hz, and $A_1$9 overall, while neural network evaluation costs $A_2$0 ms per object per view on an NVIDIA RTX 2080 Ti (Tang et al., 2020). The paper does not explicitly use the term ART, but it describes a learned, dynamic, object-level transfer operator that is ART-like in both basis representation and runtime use.

NAT pushes this operator view further. It replaces tabulated precomputed acoustic transfer data with an implicit neural representation,

$A_2$1

where $A_2$2 is a scene-condition vector encoding geometry, material, and pose parameters. Training data come from CUDA-accelerated BEM or Monte-Carlo-based BEM approximation under smooth Neumann conditions. The final model is small—about $A_2$3 to $A_2$4 MB—and inference for half-minute audio is reported at $A_2$5 to $A_2$6 ms, with numerical accuracy and runtime efficiency within several milliseconds for 30s audio (Jin et al., 6 Jun 2025). A plausible implication is that NAT functions as a neural acoustic transfer operator over a continuous parameter space rather than over a finite catalogue of precomputed ART states.

DART remains closest to classical ART. It factorizes the discrete kernel into geometry- and material-dependent components,

$A_2$7

and makes the material matrix $A_2$8, reflection coefficients $A_2$9, and, in the parametric version, BSDF mixture weights $$P(A_1, A_2, t) = \iint_{A_2} \iint_{A_1} L\big(x, v(x|x'), t\big)\; d\omega_x(x')\; dA_{\text{p}, x'}(x).$$0, differentiable. It optimizes these parameters directly against measured echograms by automatic differentiation through sparse ART recursions. In sparse measurement scenarios, DART exhibits favorable properties and better generalization than existing signal processing and neural network baselines while remaining a simple, fully interpretable system (Lee et al., 19 Sep 2025).

5. Delay-network and modal approximations

A separate line of work treats ART as a physically grounded template for low-cost reverberation structures. “Room Acoustic Rendering Networks with Control of Scattering and Early Reflections” states explicitly that recent work introduced delay network designs that approximate ART, a GA model that simulates the transfer of acoustic energy between discrete surface patches in an environment (Scerbo et al., 2023). The key structural equivalence is between ART’s patch-to-patch recursion and a Feedback Delay Network (FDN): ART radiance becomes the delay-line state, ART delays become line lengths, and the ART scattering matrix becomes the feedback matrix after a square-root amplitude mapping.

This viewpoint yields concrete algorithmic consequences. The ART scattering matrix is sparse and block-structured because a path ending at a patch can only be scattered along paths attached to that patch. The paper therefore proposes physically based, stability-preserving feedback matrix constructions. One design uses Sinkhorn–Knopp scaling to produce a doubly stochastic target matrix; another uses a uniform matrix controlled by a per-patch scattering coefficient $$P(A_1, A_2, t) = \iint_{A_2} \iint_{A_1} L\big(x, v(x|x'), t\big)\; d\omega_x(x')\; dA_{\text{p}, x'}(x).$$1, with specular entries $$P(A_1, A_2, t) = \iint_{A_2} \iint_{A_1} L\big(x, v(x|x'), t\big)\; d\omega_x(x')\; dA_{\text{p}, x'}(x).$$2 and diffuse entries $$P(A_1, A_2, t) = \iint_{A_2} \iint_{A_1} L\big(x, v(x|x'), t\big)\; d\omega_x(x')\; dA_{\text{p}, x'}(x).$$3. Both are then projected to a sign-agnostic unilossless matrix. A second extension allows an arbitrary number of early reflections to be modeled with high accuracy through high-order injection operators, so the network can be scaled between computational cost and early reverb precision (Scerbo et al., 2023). This suggests that ART is not only a simulation method but also a design prior for reverberation networks.

MoD-ART recasts the same ART system through modal decomposition. Starting from Time-Domain ART as a recursive discrete-time LTI system, it decomposes the echogram transfer function into poles and residues: $$P(A_1, A_2, t) = \iint_{A_2} \iint_{A_1} L\big(x, v(x|x'), t\big)\; d\omega_x(x')\; dA_{\text{p}, x'}(x).$$4 Each pole is an energy decay mode. The dominant real positive eigenvalue corresponds to the global late decay slope, while additional real positive poles explain multiple decay slopes in coupled volumes; complex poles explain oscillatory energy patterns such as flutter echoes. A crucial structural result is that each mode’s positional dependence can be represented by one complex scalar per source and one per listener, while the environment-dependent poles and undriven residues are computed offline (Scerbo et al., 2024). Through a complexity analysis, MoD-ART is reported to compare very favourably with ray tracing and to capture multiple decay slopes and flutter echoes (Scerbo et al., 2024).

Both the delay-network and modal lines preserve the ART ontology: sound is carried by energy-transfer operators on a structured graph. What changes is the computational packaging—recursive network, unilossless matrix, or reduced pole-residue model.

6. Uses, scope, and limitations

ART and ART-derived models are used wherever directional, position-dependent, or late-reverberant sound fields must be updated efficiently. The literature explicitly targets VR, AR, and gaming, where room acoustic synthesis enhances listeners’ sense of immersion, realism and externalisation; interactive applications with dynamic obstacles and changing topology; realistic audio-visual generation in Nerfstudio-compatible pipelines; and acoustic field learning or material optimization from sparse measurements (Scerbo et al., 2023, Tang et al., 2020, Brunetto et al., 2024, Jin et al., 6 Jun 2025, Lee et al., 19 Sep 2025). MoD-ART further addresses scenarios with multiple sound sources and listeners in complex or coupled spaces where late reverberation depends on their positions and therefore must be adapted to movements in real time (Scerbo et al., 2024).

Several misconceptions recur. First, ART is not a full wave solver. It is a geometrical-acoustics or radiative-transfer framework centered on acoustic radiance or directional energy density, not on pressure phase. Wave phenomena such as interference and diffraction are ignored or treated with separate approximations in the core geometric formulation (Lee et al., 19 Sep 2025). Second, neural audio-visual or neural transfer systems do not automatically implement classical ART equations. NeRAF is described as a learned neural mapping from 3D scene geometry and materials to a spatially varying acoustic field and as an implicit, data-driven version of radiance or ray transfer for sound, not as a direct solver of the classical ART equations (Brunetto et al., 2024). NAT likewise sits squarely in the ART tradition but replaces explicit tabulated transfer data with an implicit neural representation (Jin et al., 6 Jun 2025). Third, reflective boundary conditions do not exhaust all boundary physics: radiative-transfer analyses of half-spaces and slabs identify one-wavelength boundary layers and coherent weak-localization effects that standard geometric ART does not represent (Messaoudi et al., 2022, Messaoudi et al., 2023).

The limitations are correspondingly domain-specific. Classical and differentiable ART inherit the high-frequency assumptions of geometric acoustics and therefore weaken at low frequencies or in strongly diffractive environments (Lee et al., 19 Sep 2025). Delay-network approximations remain approximations of ART, not of the full wave equation (Scerbo et al., 2023). NeRAF assumes static scenes and a single sound source at a time, is scene-specific, and does not model frequency-dependent reflection or absorption explicitly (Brunetto et al., 2024). NAT focuses on local radiation, scattering, diffraction, and cavity effects around objects rather than full room acoustics with diffuse reverberation and long tails, and its reported experiments consider magnitude $$P(A_1, A_2, t) = \iint_{A_2} \iint_{A_1} L\big(x, v(x|x'), t\big)\; d\omega_x(x')\; dA_{\text{p}, x'}(x).$$5 only (Jin et al., 6 Jun 2025). The learned acoustic scattering-field model is trained and validated only up to 1000 Hz for its wave-based low-frequency component and assumes sound-hard objects (Tang et al., 2020). Beyond Helmholtz, radiative-transfer theory shows that omitting operator contributions from density fluctuations can significantly bias transport parameters even when disorder is weak (Baydoun et al., 2016).

Taken together, these strands define ART less as a single algorithm than as a research program. In its classical form, ART discretizes the acoustic rendering equation into radiance exchange between patches and directions. In its transport-theoretic form, it is a kinetic description derived from the Wigner transform of the wave field. In its modern variants, it becomes a differentiable, modal, delay-network, or neural operator. The common denominator is the treatment of sound propagation as transfer of directional acoustic energy through geometry, materials, and time.