Tensor Networks for Block Adaptive Thermal Radiation Transport

arXiv:2611.08410 · astro-ph.IM · Submitted 2026-11-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.

Vera: We're looking at the paper: "Tensor Networks for Block Adaptive Thermal Radiation Transport".

Jocelyn: Here's the gist of the paper. Tensor networks for block adaptive thermal radiation transport introduces a framework combining block-adaptive mesh refinement with block-local tensor-train representations of photon specific intensity to mitigate the curse of dimensionality in high-dimensional thermal radiation transport (Page 1).

Vera: This approach allows fine-scale features to be resolved while limiting the impact of locally complex structure on tensor ranks elsewhere in the domain (Page 1).

Jocelyn: Governing Equations and Discretization The paper begins by defining the governing equations for thermal radiation transport, which involve the specific intensity If, material temperature T, and a multigroup frequency domain discretization (Page 3).

Vera: The numerical setup involves a two-dimensional spatial domain discretized into uniform cells with centers Xˆ and an angular space discretized into Nθ × Nϕ angles (Page 4).

Title and authors: Vera: So we're talking about how they name this work, Tensor Networks for Block Adaptive Thermal Radiation Transport, and what that actually means for us in the observational side of things.

Jocelyn: It sounds like they are trying to make a method where you can get high angular resolution, which is important for studying anisotropy and shadowing in radiation fields.

Subrahmanyan: From a theoretical standpoint, if they can manage the rank growth caused by localized nonseparable structure, that implies there's a way to handle complex fluid dynamics coupled with radiation transport on large scales without brute force calculation.

Vera: The core idea is using these tensor networks to compress the specific intensity data into something much smaller than a standard grid representation would require.

Jocelyn: And then they couple that compression with block adaptive mesh refinement so they can focus their computational power where the physics is really demanding, like at sharp material interfaces or shocks.

The paper's summary: Vera: They summarize it by saying that the photon specific intensity depends on space, angle, and frequency all together, which creates a massive state space whose cost grows very fast with resolution.

Jocelyn: So the solution they propose involves using tensor network methods to mitigate this curse of dimensionality when the solution has some low-rank structure.

Subrahmanyan: They are specifically pointing out that localized nonseparable structures can cause substantial rank growth in these tensor networks, which is the main bottleneck they are trying to manage.

Vera: And they introduce block-structured adaptive mesh refinement as a way to concentrate the resolution exactly where those high-rank regions are located, while keeping the global calculation manageable.

Jocelyn: They also discuss how they use a splitting based scheme with an explicit transport operator and an implicit emission absorption scattering operator to actually solve the equations numerically.

The paper's improvements: Vera: The paper highlights some key improvements, like using a block-local tensor-train representation that splits the intensity into three cores: space, theta, and phi.

Jocelyn: This splitting reduces the storage requirement from something that grows with all those dimensions to something much more manageable based on the number of blocks.

Subrahmanyan: They focus on making sure this representation is maintained while still allowing for conservative block communication, which is crucial for maintaining accuracy across different parts of the simulation.

Vera: The block adaptive mesh refinement part only changes the spatial discretization locally within a block, leaving the angular cores unchanged, which is a specific way they handle that.

Jocelyn: They use rounding strategies by allocating an error budget between finite volume spatial error and rounding error to decide when to refine or coarsen a block based on an indicator called eta b.

Conclusion: Vera: So to wrap up, the main point of Tensor Networks for Block Adaptive Thermal Radiation Transport is that they successfully integrate tensor train methods into block structured adaptive mesh refinement frameworks.

Jocelyn: They show that localizing high rank structure to blocks that need it actually transforms the rank distribution into a spatially adaptive resource, which is really clever.

Subrahmanyan: It means we can achieve compression factors exceeding ten to the power of three on problems where uniform grid tensor methods fail because they require too much global rank <ref:2611.08410#pg1>.

Vera: It provides a scalable path toward high fidelity multiphysics simulation by matching the computational cost directly to how difficult the local solution is.

Jocelyn: The paper shows that AMR and tensor compression work together: AMR reduces spatial degrees of freedom, and the block-local tensor factorization reduces the rank burden away from regions needing high angular fidelity.

Subrahmanyan: It confirms that this approach lets you handle complex radiation transport on a multiscale mesh efficiently by isolating the hard parts.

Alex A. Gorodetsky, Patrick D. Mullen, Aditya S. Deshpande, Joshua C. Dolence, Chad D. Meyer, Jonah M. Miller, Luke F Roberts

Department of Aerospace Engineering, University of Michigan · Michigan SPARC, Los Alamos National Laboratory

astro-ph.IM

Submitted: 2026-11-05

Updated: 2026-11-05

Comments: 34 pages, 15 figures, LA-UR-26-24582

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: Tensor networks for block adaptive thermal radiation transport introduces a framework combining block-adaptive mesh refinement with block-local tensor-train representations of photon specific…

Key concepts

Tensor Train (TT) Representation
This is a mathematical method used to compress large matrices, like the specific intensity of radiation. Instead of storing all possible values globally, it breaks the data down into smaller 'cores' that are linked together in a chain. This drastically reduces storage needs when dealing with high-dimensional problems.
Block Adaptive Mesh Refinement (AMR)
This technique dynamically adjusts the spatial grid resolution only where it is needed. Instead of refining the entire simulation domain uniformly, AMR focuses refinement on specific blocks that contain complex features, saving computational effort while maintaining accuracy in critical areas.
Block-Local Tensor-Train
Instead of one massive tensor for the whole simulation, this method uses separate TT representations for different spatial blocks. This isolates high-rank (complex) structures to specific regions, preventing them from inflating the rank across the entire simulation domain.

Terminology

Summary

Tensor networks for block adaptive thermal radiation transport introduces a framework combining block-adaptive mesh refinement with block-local tensor-train representations of photon specific intensity to mitigate the curse of dimensionality in high-dimensional thermal radiation transport (Page 1). p. 0 This approach allows fine-scale features to be resolved while limiting the impact of locally complex structure on tensor ranks elsewhere in the domain (Page 1). p. 0

Governing Equations and Discretization

The paper begins by defining the governing equations for thermal radiation transport, which involve the specific intensity If, material temperature T, and a multigroup frequency domain discretization (Page 3). The numerical setup involves a two-dimensional spatial domain discretized into uniform cells with centers Xˆ and an angular space discretized into Nθ × Nϕ angles (Page 4). p. 2 A standard ghost-cell based approach is used to enforce boundary conditions by extending the domain to Xˆext and applying a boundary condition operator Bm (Page 4). p. 2 The transport solver is then formulated using a splitting-based scheme involving the explicit transport operator T and the spatially local implicit emission, absorption, and scattering operator S (Page 4). p. 2

Block-Local Tensor-Train Representation

The specific intensity in each block is represented using a tensor-train (TT) ansatz, split into three cores: space, θ, and ϕ (Page 6). This representation reduces the storage requirement from NxNθNϕ to r1bNx+r1b r2bNθ+r2b Nϕ (Page 6). p. 5 The most important TT operation required for ghost cell communication is addition, which yields a block-diagonal structure for the central core (Page 6). p. 5 This additive structure must be controlled through a rounding procedure such as TT-SVD to manage rank growth (Page 6). p. 5

Block Adaptive Mesh Refinement and Rounding

The core contribution is the block-local AMR procedure that preserves the TT representation while maintaining conservative block communication (Page 7). The process involves filling ghost data using neighbor boundary operators Bm c→b, and computing flux correction on shared interfaces using the operator C x c→b (Page 7). p. 3 Adaptive refinement changes only the spatial discretization local to a block, acting only on the spatial core while leaving angular cores unchanged (Page 8). p. 7 Rounding strategies are applied by allocating an error budget between finite-volume spatial error and rounding error, leading to criteria like Equation (A21) for refinement based on the indicator ηb exceeding a threshold (Page 30). p. 7

Compression Metrics and Performance

The paper introduces the diagnostic Cd, defined as Sglobal-TT / Sblock-TT, to quantify the storage benefit of block-local tensor representation relative to a single global TT representation (Page 9). Results show that for rank-1 problems, Cd decreases with Nb because angular storage overhead dominates when NθNϕ is large (Page 10). p. 8 For the 2D line source problem, TT-AMR achieves a sweet spot near Nb ∼ 100 where Cd reaches ∼5 for TT-DD and ∼6 for TT-AMR, demonstrating that block decomposition isolates high-rank structure to localized regions (Page 13). p. 0

Error Balancing and Practical Considerations

The error model decomposes the global error into finite-volume spatial error and rounding error, allowing the adaptive algorithm to balance these contributions through choices like intensity-weighted partitioning (Page 30). The sparse block strategy allows blocks with low Frobenius norm Fb < Ftol to be assigned infinite rounding tolerance, suppressing numerical artifacts from superluminal signals (Page 21). p. 7 This approach demonstrates that AMR and tensor compression are mutually reinforcing, as AMR reduces spatial degrees of freedom while the block-local TT representation reduces the rank burden away from difficult regions (Page 25). p. 0

Conclusion

The work successfully integrates tensor train methods into block-structured AMR frameworks, demonstrating that localizing high-rank structure to blocks that need it transforms the rank distribution into a spatially adaptive resource (Page 25). Compression factors exceeding 10 cubed are achievable on problems where uniform-grid TT methods fail due to excessive global rank requirements (Page 26). p. 0 The framework provides a scalable path toward high-fidelity multiphysics simulation by matching computational cost to local solution difficulty (Page 25). p. 0

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Block-local tensor-train representation for thermal radiation transport allows for localized low-rank compression, achieving compression factors exceeding 103 on multiscale, adaptively refined meshes by confining nonseparable structure to specific blocks.

  2. The framework enables a synergistic approach where adaptive mesh refinement reduces the number of spatial degrees of freedom while the block-local tensor factorization reduces the rank burden away from regions requiring high angular fidelity.

  3. A rank-based refinement criterion can be used to refine blocks based on their local tensor ranks, addressing the mismatch between J-based refinement (energy concentration) and angular complexity (nonseparable structure).

  4. The interlinked adaptive strategy distributes a prescribed per-timestep error budget between spatial discretization error and rounding error, using indicators like eta b to trigger refinement or coarsening dynamically.

  5. The sparse block strategy can be employed by assigning infinite rounding tolerance to blocks with low Frobenius norms, effectively forcing them to rank one and suppressing spurious rank growth from numerical artifacts in quiescent regions.

  6. The method allows for the simultaneous reduction of spatial degrees of freedom and tensor rank, achieving compression factors exceeding 103 on problems where uniform-grid tensor methods might fail due to excessive global rank requirements.

Sources

Related papers