MD-PNOP: Equation-Recast Neural Operators for Minimal-Data Extrapolation and PDE Solver Acceleration

Technical Summary: MD-PNOP – Equation-Recast Neural Operators for Minimal-Data Extrapolation and PDE Solver Acceleration

Problem Statement
Traditional numerical solvers for partial differential equations (PDEs), such as finite element or finite difference methods, face significant computational bottlenecks in large-scale parametric studies and design optimization. While data-driven approaches like Physics-Informed Neural Networks (PINNs) and Neural Operators (e.g., DeepONet, FNO) offer faster inference, they suffer from critical limitations:

1. Data Scarcity and Cost: Training robust surrogates typically requires extensive datasets spanning the entire parameter space, which is often prohibitively expensive to generate.
2. Extrapolation Failure: Neural networks generally struggle to generalize to parameter regimes unseen during training, leading to accuracy degradation.
3. Physical Consistency: Purely data-driven models often fail to strictly enforce governing equations, raising concerns about reliability and interpretability in safety-critical applications (e.g., nuclear reactor design).
4. Black-Box Nature: The lack of strict physical constraints in standard training pipelines limits trust in surrogate predictions for decision-making.

Methodology: The MD-PNOP Framework
The authors propose Minimal-Data Parametric Neural Operator Preconditioning (MD-PNOP), a hybrid framework that integrates neural operators into traditional model-based iterative solvers. The core innovation lies in an equation-recast strategy inspired by perturbation theory, which enables a neural operator trained on a single parameter configuration to solve problems with arbitrary, unseen parameter settings without retraining.

The methodology proceeds in three stages:

1. Equation Recast (Parameter Generalization):
   Instead of training a neural operator to map parameters directly to solutions, the governing equation for a new parameter set $\mathbf{p}$ is reformulated relative to a nominal training parameter set $\mathbf{p}^*$. The parameter difference $\Delta\mathbf{p} = \mathbf{p} - \mathbf{p}^*$ is treated as a perturbation. The original operator $\mathcal{L}(\mathbf{p})$ is decomposed into the trained operator $\mathcal{L}(\mathbf{p}^*)$ and an operator difference term $\delta\mathcal{L}(\Delta\mathbf{p})$.
   The equation is recast as:
   $$\mathcal{L}(\mathbf{p}^*)[u(\mathbf{x})] = S(\mathbf{x}) - \delta\mathcal{L}(\Delta\mathbf{p})[u(\mathbf{x})]$$
   Here, the parameter deviation is absorbed into a modified source term on the right-hand side. This allows the use of a neural operator trained only on $\mathbf{p}^*$ to iteratively solve for $u(\mathbf{x})$ under new parameters $\mathbf{p}$ by treating the operator difference as an additional source.

2. Neural Operator Preconditioning (Acceleration):
   The recast equation is solved iteratively using the pretrained neural operator $\mathcal{L}_{NO}^{-1}(\mathbf{p}^*)$ as a fast predictor:
   $$u^{(k+1)}(\mathbf{x}) = \mathcal{L}_{NO}^{-1}(\mathbf{p}^*) \left[ S(\mathbf{x}) - \delta\mathcal{L}(\Delta\mathbf{p})[u^{(k)}(\mathbf{x})] \right]$$
   This iterative process generates a high-quality initial guess ($u_{NO}$) that is close to the true solution.

3. Hybrid Solver Integration (Accuracy Guarantee):
   The neural operator prediction $u_{NO}$ is not used as the final solution. Instead, it serves as the initial guess for a traditional, high-fidelity model-based solver (e.g., $S_N$ method for transport equations). The model-based solver then performs a few iterations to converge to the exact solution satisfying the original governing equations.

   • Simple Preconditioning (SP): The neural operator replaces the inner solver entirely for the iterative process, with a final correction by the model-based solver.
   • Constrained Preconditioning (CP): The model-based solver is invoked at every inner iteration step to constrain the neural operator's prediction, enhancing stability for complex or heterogeneous problems.

Key Contributions

• Minimal-Data Training: The framework requires training data from only a single set of constant parameters, drastically reducing data acquisition costs compared to standard neural operator approaches.
• Systematic Extrapolation: By recasting parameter variations as source terms, the method achieves systematic generalization to heterogeneous, sinusoidal, and discontinuous parameter distributions without retraining.
• Physics-Consistent Accuracy: The final solution is strictly generated by the model-based solver, ensuring full-order fidelity and strict adherence to physical laws, thereby eliminating "black-box" concerns.
• Architecture Agnosticism: The framework is demonstrated using both Deep Operator Networks (DeepONet) and Fourier Neural Operators (FNO), proving its independence from specific neural operator architectures.

Results
The framework was evaluated on the one-dimensional Boltzmann transport equation (neutron transport) across three problem types:

1. Fixed Source Problems: Tested with constant, heterogeneous (sinusoidal), and discontinuous cross-sections, including anisotropic scattering terms not present in training.
2. Single-Group Eigenvalue Problems: Tested on constant, sinusoidal, and step-function cross-section profiles, including high-dominance ratio systems.
3. Multigroup Coupled Eigenvalue Problems: Tested on a three-group system with tightly coupled scattering and fission terms.

Performance Metrics:

• Accuracy: In all test cases, the hybrid solvers (MD-PNOP) achieved $L_2$-norm errors on the order of $10^{-4}$, matching the benchmark model-based solver. Standalone neural operators exhibited significantly higher errors ($10^{-3}$ to $10^{-2}$) when extrapolating to unseen parameters.
• Computational Efficiency: MD-PNOP consistently reduced total computational time by approximately 50% (ranging from ~35% to ~60% depending on the case and architecture) compared to the standalone model-based solver.
  • For simple, homogeneous problems, DeepONet-based preconditioning offered the highest speedup (up to 97% reduction in time for eigenvalue cases).
  • For complex, heterogeneous problems, FNO-based preconditioning demonstrated superior generalization and stability, often outperforming DeepONet in accuracy while maintaining significant speedups.

Significance and Claims
The paper claims that MD-PNOP establishes a robust strategy for accelerating PDE solvers by balancing computational efficiency with physical accuracy. Its significance lies in:

• Overcoming the Data Bottleneck: It enables the use of neural operators in scenarios where generating large training datasets is infeasible.
• Bridging the Gap: It effectively bridges the gap between the speed of data-driven inference and the reliability of model-based solvers.
• Generalizability: The equation-recast formulation is defined at the operator level, making the framework applicable to a wide range of PDEs and parameter variations without requiring architecture-specific tuning.
• Safety and Reliability: By ensuring the final solution is derived from a traditional solver, the framework addresses interpretability and safety concerns, making it suitable for critical engineering applications like nuclear reactor design and optimization.

The authors emphasize that while the framework introduces additional computation to evaluate the operator difference (source terms), the replacement of expensive spatial sweeps with fast neural operator inference results in a net reduction in wall-clock time. The approach is presented not as a replacement for high-fidelity solvers, but as a powerful preconditioning tool to accelerate them.