pcNN-SN: a physics-enhanced predictor-corrector neural network algorithm coupled with the discrete ordinates method for neutron transport equations

arXiv:2610.20619 · physics.comp-ph, cs.LG · Submitted 2026-09-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "pcNN-SN: a physics-enhanced predictor-corrector neural network algorithm coupled with the discrete ordinates method for neutron transport equations".

Jane: The gist The proposed algorithm improves coarse-grid solutions for neutron transport calculations by coupling a physics-enhanced predictor-corrector neural network with the discrete ordinates method,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we've been looking at the "pcNN-SN: a physics-enhanced predictor-corrector neural network algorithm coupled with the discrete ordinates method for neutron transport equations" paper, which is Cong Liua, Junxia Weia, Shuai Wanga, Heng Yong’s work.

Jane: It really boils down to using a physics-informed neural network structure to predict how much you need to correct a coarse grid solution based on higher accuracy data and local physics features.

Lu: The core contribution is the predictor-corrector neural network design that uses dimensionless physical quantities, which reduces the input space dimensionality significantly, making it tractable for training.

Meng: From my side, I'm looking at how this could be used in production—if it works as advertised and saves sixty percent of compute time over mesh refinement, that’s a huge win for deployment cost.

Tom: The paper shows that this method can achieve accuracy levels comparable to fine-grid results on coarse grids, which is the main claim they are making about its performance.

Jane: It means that instead of spending all our resources on hyper-fine meshes, we could use a coarser grid and let the AI handle the necessary corrections efficiently.

Tom: The authors say this physics-based network design allows for a certain degree of reusability under different computational conditions, which is important for practical use.

Lu: And they also address that by including specific source reconstruction algorithms compatible with the SN framework to get higher-order source information.

Jane: So, in simple terms, they’re proposing a way to use machine learning to relax the requirements on how fine your simulation grid needs to be for neutron transport calculations.

Conclusion: Tom: So we've looked at this paper, "pcNN-SN," which uses a physics-enhanced neural network to fix coarse grid solutions for neutron transport calculations and achieves accuracy near fine grid levels while cutting computation time by over sixty percent.

Jane: It sounds like they’re taking something that’s traditionally very computationally expensive—those high-dimensional features in transport simulations—and applying machine learning to make the coarse grids actually work well.

Lu: Exactly. The core idea is this predictor-corrector neural network, or pcNN, which learns how to predict those necessary correction factors by looking at local physical features directly from the transport equations themselves.

Meng: From a production standpoint, that sixty percent time saving is what matters most to me. If we can run simulations much faster without sacrificing accuracy, that’s a huge win for anything involving large-scale AI training or design cycles.

Tom: And the authors show they've built this on high- and low-accuracy data, using dimensionless physical quantities to keep the input space manageable for the network training.

Jane: So, what does this mean practically? It suggests that we might relax those strict requirements on grid spacing in traditional methods like discrete ordinates because a smart correction model can compensate for some of that coarseness.

Tom: Right. The conclusion is pretty straightforward—the pcNN-SN results on coarse grids still agree well with the high-accuracy solutions, and the cost to evaluate this network is less than what you'd spend just refining the mesh itself.

Jane: That’s a significant shift in how we think about solving these complex physics problems; it moves us toward using AI as an intelligent post-processor rather than just a replacement for brute force computation.

Lu: And they stress that this physical design allows the model to be quite reusable under different conditions, which is key because we can't just train one model and expect it to work everywhere in a strongly inhomogeneous medium.

Meng: I’m still worried about the "strongly inhomogeneous media" part; if the physics gets really weird, does this AI correction still hold up without massive retraining?

Tom: That’s the next big question, right? Because they’ve shown this on simple three-D problems, but moving to more complex real-world scenarios is where we need to see how robust this whole framework really is.

Institute of Applied Physics and Computational Mathematics · National Key Laboratory of Computational Physics

physics.comp-ph, cs.LG

Submitted: 2026-09-02

Updated: 2026-09-02

Comments: 35 pages, 17 figures

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

Importance score: 81/100

The gist: The gist The proposed algorithm improves coarse-grid solutions for neutron transport calculations by coupling a physics-enhanced predictor-corrector neural network with the discrete ordinates method,

Key concepts

Predictor-Corrector Neural Network (pcNN)
This is a machine learning model trained using high- and low-accuracy data. It is designed to predict necessary correction factors for coarse-grid solutions during the transport calculation process, effectively bridging the gap between coarse and fine grid results.
Discrete Ordinates Method (SN)
SN is a standard numerical technique used to solve neutron transport equations. The proposed method integrates the pcNN directly into this sweep process, allowing the network to provide online corrections as the simulation moves through different spatial points.
Physics-Enhanced Inputs
The network inputs are derived from local physical features extracted directly from transport equations, such as subgrid-scale information. Dimensionless transformations are applied to these quantities to simplify the input space and improve the network's training efficiency.
Conservation Constraints
Hard and soft constraints are built into the neural network model. These constraints ensure that even when correcting coarse solutions, fundamental physical properties like particle balance equations remain satisfied, guaranteeing the corrected solution retains its original conservation properties.

Terminology

Summary

The gist The proposed algorithm improves coarse-grid solutions for neutron transport calculations by coupling a physics-enhanced predictor-corrector neural network with the discrete ordinates method, achieving accuracy comparable to fine-grid results and reducing computation time by more than 60%

How it works

The core of the method is a predictor-corrector neural network (pcNN) designed to predict correction factors for coarse-grid solutions based on high- and low-accuracy data The network is informed by local physical features containing subgrid-scale information selected from transport equations Dimensionless transformation of physical quantities is used to reduce input space dimensionality and narrow parameter distributions, which facilitates network training Physical constraints are incorporated into the network model to ensure that the corrected solution retains its original conservation properties

Physical Feature Inputs for pcNN

The paper emphasizes the necessity of incorporating higher-order source information because grid-averaged source information does not account for variations in the source distribution within a coarse grid A source reconstruction algorithm is developed compatible with the SN source iteration and sweep framework to reconstruct higher-order source information based on the flux solution from previous iterations This involves reconstructing the flux distribution using interpolation basis functions and deriving the first-order moment of the scattering source, which is defined as Qx s x =   (20)

Loss Functions and Constraints

To ensure that the corrected numerical solution preserves inherent conservation properties, hard and soft constraints are introduced into the network model The loss function is constructed to predict outgoing angle flux correction factors, and a hard constraint approach is used to satisfy the particle balance equation by fixing the grid-averaged value Soft constraints are added based on the balance relationship, where an equivalent correction factor for the grid-averaged solution is calculated to ensure that if fix ψ approaches the fine-grid reference value, then fix ave f similarly approaches the training data ave n f

Coupling Framework and Verification

The pcNN model is internally coupled with the traditional SN sweep calculation process, where local features of the angular flux solution are fed into the neural network to generate correction factors online Training data generation involves projecting fine-grid physical quantities onto coarse grids to obtain equivalent physical parameters for local low-accuracy angular flux calculations Numerical verification on 3-D single-medium problems showed that the pcNN-SN algorithm yields error norms significantly lower than traditional SN results obtained after halving the grid size, and it achieves overall computational time savings exceeding 60%

Reusability and Performance

The dimensionless treatment of physical quantities is crucial because it reduces the dimensionality of the network inputs and narrows the range of values, which facilitates reusability across different cross section and grid conditions Data augmentation, including applying rotational transformations to both inputs and outputs, helps guarantee permutation-equivariant properties for orthogonal grids by expanding the training dataset into six groups The algorithm demonstrates good reusability under various conditions because it captures the correspondence between high- and low-accuracy solutions within the parameter range

Conclusion

The study concludes that by coupling machine learning with traditional SN transport simulation, a novel physics-enhanced approach can relax requirements for grid spacing in the SN difference method, and the pcNN-SN results on coarse grids still agree well with high-accuracy solutions The network evaluation cost is lower than the cost introduced by mesh refinement, and the pcNN-SN algorithm achieves overall computational time savings exceeding 60% This study explores a method coupling machine learning with traditional SN transport simulation, and we apply it to simple geometry problems Further research is needed to develop a unified network model for strongly inhomogeneous media in complex problems The output of our network can incorporate correction factors corresponding to subgrids to enhance computational resolution

--- Page 35 ---

The gist The proposed algorithm improves coarse-grid solutions for neutron transport calculations by coupling a physics-enhanced predictor-corrector neural network with the discrete ordinates method, achieving accuracy comparable to fine-grid results and reducing computation time by more than 60%

Conclusion

The study concludes that by coupling machine learning with traditional SN transport simulation, a novel physics-enhanced approach can relax requirements for grid spacing in the SN difference method, and the pcNN-SN results on coarse grids still agree well with high-accuracy solutions The network evaluation cost is lower than the cost introduced by mesh refinement, and the pcNN-SN algorithm achieves overall computational time savings exceeding 60% This study explores a method coupling machine learning with traditional SN transport simulation, and we apply it to simple geometry problems Further research is needed to develop a unified network model for strongly

Improvements for AI systems

  1. The improved system can achieve a level of accuracy comparable to that of fine-grid results for coarse-grid solutions, as demonstrated by the pcNN-SN algorithm effectively correcting low-accuracy solutions.

  2. The system will reduce computation time for three-dimensional test problems by more than 60%, leveraging the physics-based network design and internal coupling of pcNN and SN transport sweeps.

  3. The model can be made highly reusable across different computational conditions because it fully incorporates physical characteristics and the network model demonstrates excellent reusability across different cross section and grid conditions.

  4. The system can handle high-dimensional input spaces more effectively by applying dimensionless transformations, which successfully reduce the dimensionality of the network’s input space and narrow the range of parameter distributions, thus facilitating network training.

  5. The system will strictly preserve original particle conservation properties because it incorporates hard and soft constraints into the network model to ensure that corrected solutions satisfy Eq. (24), which represents the balance between inflow, outflow, collision, and generation within the grid.

  6. The AI system can perform online correction of the local numerical solution by coupling pcNN with SN transport sweeps, where the neural network model performs online correction of the local numerical solution based on physical field features obtained from SN sweeps as input.

  7. The system can generalize better to unseen scenarios by adopting a data augmentation approach, which expand[s] the training dataset by applying rotational transformations to both inputs and outputs to satisfy permutation-equivariant constraints.

Related papers