Abstract
Tensor network techniques, known for their low-rank approximation ability that breaks the curse of dimensionality, are emerging as a foundation of new mathematical methods for ultra-fast numerical solutions of high-dimensional Partial Differential Equations (PDEs). Here, we present a mixed Tensor Train (TT)/Quantized Tensor Train (QTT) approach for the numerical solution of time-independent Boltzmann Neutron Transport equations (BNTEs) in Cartesian geometry. Discretizing a realistic three-dimensional (3D) BNTE by (i) diamond differencing, (ii) multigroup-in-energy, and (iii) discrete ordinate collocation leads to huge generalized eigenvalue problems that generally require a matrix-free approach and large computer clusters. Starting from this discretization, we construct a TT representation of the PDE fields and discrete operators, followed by a QTT representation of the TT cores and solving the tensorized generalized eigenvalue problem in a fixed-point scheme with tensor network optimization techniques. We validate our approach by applying it to two realistic examples of 3D neutron transport problems, currently solved by the PARallel TIme-dependent SN (PARTISN) solver. We demonstrate that our TT/QTT method, executed on a standard desktop computer, leads to a yottabyte compression of the memory storage, and more than 7500 times speedup with a discrepancy of less than 1e-5 when compared to the PARTISN solution.
Introduction
In nuclear engineering, simulating how neutrons move through a medium is critical for designing reactors and ensuring safety. This is governed by the Boltzmann Neutron Transport Equation (BNTE), an integro-differential equation that lives in a high-dimensional space: three spatial dimensions, two angular dimensions, and one energy dimension.
When discretizing this 6D problem using standard methods, researchers hit the "curse of dimensionality." The number of grid points grows exponentially, leading to systems of equations so massive they require high-performance computing (HPC) clusters and "matrix-free" techniques just to fit in memory. Even then, a single simulation can be computationally exhausting. This paper presents a framework to "break" this curse by fundamentally changing how the data is represented and processed using tensor networks.
Methods
The researchers introduced a framework that combines traditional discretization with cutting-edge tensor algebra:
- Standard Discretization: The process begins with "diamond differencing" for space, "multigroup" for energy, and "discrete ordinates" for angles. This transforms the PDE into a massive generalized eigenvalue problem.
- Tensor Train (TT) Representation: Instead of storing the solution as a giant 6D array, it is represented as a "Tensor Train"—essentially a chain of small, low-rank tensors. This drastically reduces storage requirements.
- Quantized Tensor Train (QTT): To push compression even further, the authors apply "quantization." This involves reshaping the TT cores themselves into even higher-dimensional tensors of size 2, exploiting the underlying logarithmic structure of the operators.
- Fixed-Point Solver: A fixed-point iteration scheme was redesigned to work entirely within the TT/QTT format using optimization techniques like the Alternating Minimal Energy (AMEn) method. This allows for solving the largest eigenvalue (criticality) without ever "unpacking" the full, massive tensor.
Key Results and Performance
The method was validated using two primary benchmarks: a 1D slab case study for mathematical verification and a realistic 3D "Jezebel" cube variant composed of plutonium and gallium isotopes. The performance was compared against PARTISN, the industry-standard parallel transport code used at Los Alamos National Laboratory.
- Staggering Speedup: For the largest 3D benchmark (512 grid points per spatial dimension), the TT/QTT method was over 7500 times faster than the PARTISN solver when accounting for core-hours.
- Extreme Compression: The operators involved in the 3D BNTE are theoretically so large that, if fully expanded, they would require yottabytes ($10^{24}$ bytes) of storage. The TT/QTT format compresses these into a few megabytes. Specifically, the memory footprint for the solution eigenvector was reduced from ~0.5 Terabytes to just 1 Megabyte.
- High Accuracy: Despite the massive compression, the discrepancy between the TT/QTT results and the reference PARTISN solution was less than $10^{-5}$, proving that the low-rank approximation maintains the physical fidelity required for engineering.
Conclusion and Future Directions
The most striking insight of this research is the "Blessing of Dimensionality": while high dimensions usually break algorithms, the QTT format actually thrives on them. By using the right data structure, a standard desktop (10-core i9) can outperform a massive MPI-based cluster (using up to 1152 cores) by orders of magnitude.
Future work will focus on extending this approach to time-dependent problems and more complex, non-rectilinear geometries. This could eventually open the door for real-time reactor monitoring and simulation on commodity hardware, moving away from the necessity of massive supercomputer clusters for routine nuclear engineering tasks.
Summary updated