Deep-learning multi-antiderivative framework enables efficient anisotropic neutron transport solutions

A deep-learning–based numerical method has been accepted for publication that tackles one of neutron transport's tougher nuts: anisotropic scattering. The paper introduces a multi-antiderivative transformation alternating iterative framework that blends learned operators with traditional iterative solves to reduce the computational burden of anisotropic neutron transport equations.

Anisotropic scattering produces angular complexity that strains deterministic solvers and inflates compute and memory needs for high-fidelity reactor, shielding, and criticality analyses. The new framework inserts a sequence of antiderivative transformations into an alternating iterative loop, and trains neural components to approximate those transforms so the solver converges more rapidly on systems with strong angular dependence. The authors positioned this hybrid approach as a bridge between physics-based discretizations and data-driven acceleration, keeping the familiar structure of iterative transport sweeps while delegating costly operator approximations to a compact learned model.

The paper was accepted on February 11, 2026. In practical terms, the framework targets workflows where anisotropic scattering kernels make standard acceleration techniques inefficient or unstable. For reactor designers running deterministic codes, shielding engineers modeling complex angular fluxes, and students building transport experiments, the approach promises lower wall-clock times for equivalent accuracy and reduced memory pressure for high-order angular expansions. That could free weekend compute clusters and lower barriers for detailed parametric studies.

Methodologically, the multi-antiderivative transformation alternating iterative framework preserves operator splitting and iteration semantics that transport practitioners recognize. It injects neural antiderivative approximators at points in the iteration where traditional preconditioners or coarse solvers would normally act, effectively acting as a learned acceleration stage. Because the scheme remains iterative and modular, it can be slotted into existing solvers with less invasive code changes than full end-to-end machine-learned replacements.

For the Nuclear Reactions community, the immediate outcome is a new option for prototyping faster deterministic solvers without abandoning established transport infrastructure. Modelers who maintain in-house codes can experiment by training lightweight networks on representative angular distributions and integrating them as iterative boosters. Educators can use the framework as a hands-on example of hybrid physics-ML design, illustrating how targeted learning can address specific numerical bottlenecks.

Next steps will center on community validation: benchmarking against canonical anisotropic test problems, assessing robustness across energy groups and geometries, and open implementation so labs can reproduce gains on their hardware. If those follow-up studies confirm the initial promise, multi-antiderivative learned accelerators could become a practical tool in the transport toolbox, speeding iterative solves and letting practitioners focus on physics rather than prohibitive compute.