Multigroup Thermal Radiation Transport with Tensor Trains

We investigate the application of tensor train (TT) algorithms to multigroup thermal radiation transport (i.e., photon radiation transport). The TT framework enables simulations at discretizations that might otherwise be computationally infeasible on conventional hardware. We show that solutions to certain multigroup problems possess an intrinsic low-rank structure, which the TT representation leverages effectively. This enables us to solve problems where the discretized solution size exceeds a trillion parameters on a single node. We consistently achieve compression factors $>$100$\times$ and speedups $>$2$\times$. The solver is evaluated across a range of test problems with varying levels of complexity. In addition, we further analyze the low-rank structure of the merged spatio-spectral core to evaluate the potential for additional compression via more advanced TT decompositions.

💡 Research Summary

This paper presents a novel application of Tensor Train (TT) decomposition to multigroup thermal radiation transport (MRT), a problem that traditionally suffers from the curse of dimensionality due to its dependence on space, angle, and frequency. By representing the high‑dimensional solution tensor as a product of low‑rank cores—one “spatio‑spectral” core that couples spatial and frequency dimensions, and two angular cores—the authors achieve dramatic reductions in memory usage and computational cost.

The governing equations are first recast in multigroup form, discretizing the frequency domain into Nν groups while retaining angular resolution via a latitude/longitude discretization (Nθ × Nϕ). The TT representation expresses the specific intensity Iₖℓp as ˆIₖℓp = Zₖ Θ_ℓ Φ_p, where Z∈ℝ^{Nₓ·Nν × r_{zθ}} merges space and frequency, Θ∈ℝ^{r_{zθ}×Nθ×r_{θϕ}} and Φ∈ℝ^{r_{θϕ}×Nϕ} encode angular dependence, and r_{zθ}, r_{θϕ} are the TT ranks. All flux stencils (upwind, Rusanov, HLL) and source terms are reformulated as TT operations, and after each time‑step a rounding (truncation) step controls rank growth, keeping the representation compact.

The authors implement a step‑then‑truncate algorithm: (1) advance the transport operator using explicit fluxes, (2) solve the implicit source term (absorption, scattering, Planck emission) in TT form, and (3) round the resulting TT tensors to a prescribed tolerance. Compatibility of operands is ensured by padding material coefficients (opacity, scattering) with rank‑1 cores so that they share the same TT structure as the solution.

A suite of benchmark problems demonstrates the method’s efficacy. In a 2‑D hohlraum vacuum test with 128² spatial cells, 10 frequency groups, and 512 × 1024 angular directions (≈1.3 × 10¹² degrees of freedom), the TT solver attains a compression ratio C ≈ 5 × 10³ and a speed‑up S ≈ 60 relative to a conventional SN implementation. Rank evolution plots reveal that the frequency dimension remains essentially rank‑1, indicating that the multigroup solution is largely a scaled copy of the gray solution. Additional tests—including line‑driven wind models and ICF hohlraum configurations with up to 100 groups—show similar compression (C > 100) and modest speed‑ups (S ≈ 2–5). Internal rank analysis of the Z core uncovers latent low‑rank structure that could be exploited by more sophisticated tensor formats (e.g., block‑TT or tree tensor networks).

The discussion acknowledges that the current merged spatio‑spectral core does not fully separate spatial and spectral variability; in highly heterogeneous media the TT ranks may grow, limiting compression. Future work is proposed to investigate advanced decompositions, adaptive rank selection, and integration with dynamic mesh refinement. The authors also note that the choice of unit system can affect numerical conditioning, and they adopt custom units to keep solution magnitudes near unity.

In conclusion, the paper demonstrates that TT‑based solvers can make previously intractable multigroup radiation transport problems feasible on a single compute node, achieving >100× memory reduction and up to 60× runtime acceleration. This opens the door to high‑fidelity, fully multigroup simulations in astrophysics, inertial confinement fusion, and related fields, where detailed frequency dependence is essential but has been historically compromised due to computational constraints.