Kernel Learning of PDE Solution Operators

Abstract: A kernel-based approach for the learning of the solution operator of general nonhomogeneous partial differential equations (PDEs) is proposed. The method incorporates physical priors, typically encoded through the PDE operator, into a kernel ridge regression framework, and employs a regularization-based formulation to construct an operator learner. This yields a closed-form estimator that is independent of the input functions that determine the underlying PDE. From the perspective of regularization theory, the resulting estimator induces a well-defined operator that links input and output spaces, which contain the functions that define a Dirichlet problem and its solution, respectively. Consequently, it effectively shifts from a PDE solver to an operator-based solver. In contrast to standard supervised learning methods, it does not rely on paired input--output training data and enables systematic extrapolation beyond observed regimes. A full error analysis is conducted, providing convergence rates for the operator-based solver under suitable choices of regularization parameters. Extensive numerical experiments, including Darcy flow and Helmholtz equations, demonstrate that the proposed method achieves high accuracy and efficiency across a range of problem settings, and compares favorably with operator learning approaches in both approximation quality and computational cost.

• The paper introduces a kernel-based operator learning framework that leverages known PDE structure to approximate solution operators without requiring paired functional data.
• It employs an operator-theoretic formulation and Gramian regression in RKHS to yield explicit, closed-form representations with uniform convergence guarantees.
• Extensive numerical experiments demonstrate high accuracy (L² errors around 10⁻³) and significant computational savings compared to traditional methods.

Kernel-Based Operator Learning for PDE Solution Operators

Motivation and Context

The paper "Kernel Learning of PDE Solution Operators" (2605.09643) introduces an operator-theoretic kernel learning framework to approximate solution operators for general linear PDEs under Dirichlet boundary conditions. The primary motivation is to address computational efficiency and generalizability limitations in existing operator-learning and physics-informed neural network approaches, particularly for scenarios with substantial prior physical knowledge and repeated evaluations of solution maps. Contrasted against neural operator schemes like DeepONet, and the Fourier Neural Operator, the proposed method explicitly incorporates known PDE operator structure, eliminating the need for extensive paired function data and bypassing complex retraining—a significant practical advantage.

Methodological Advances

The approach models the solution space in an RKHS defined by a universal kernel (typically Gaussian), and formulates learning as an empirical risk minimization incorporating both the differential operator and boundary operator from the governing PDE as physical priors. Key technical innovations include:

• Operator-Theoretic Formulation: The solution operator $T : h \mapsto u$ is learned by minimizing a regularized squared-error loss, yielding explicit operators $T_{\Lambda,N}$ and $T_\Lambda$ for empirical and statistical settings, respectively.
• Generalized Kernel Representer Theorem: The minimizer admits a closed-form through Gramian regression involving operator-applied kernel sections. Explicit representations $u_{\Lambda,N} = \sum c_i P_{(1,0)} K(x_i, \cdot)$ are derived, with coefficients obtained as the solution to a linear system constructed from the PDE structure and sample points.
• Independence from Paired Functional Data: Unlike data-driven neural operator learning, this approach does not require paired function samples between input and output spaces. The estimator remains independent of the particular input function $h$, enabling rapid evaluation for new sources/boundary data.
• Connection to Green Functions: The kernel basis constructed can be interpreted as an approximation to the Green function associated with the PDE. This bridges classical PDE theory and modern operator learning.

Theoretical Analysis

The framework is accompanied by a comprehensive statistical error analysis:

• Estimation and Approximation Error Bound Decomposition: The total reconstruction error for the operator estimator is decomposed and characterized via pathwise concentration inequalities (Bernstein-type for self-adjoint operators) and source-regularity classes.
• Uniform Convergence Guarantees: Uniform error bounds are established over source-induced function classes, with explicit convergence rates tied to sample size $N$ and regularization parameter $\Lambda$. Specifically, the error decays as $N^{-\min\{\alpha\gamma, \frac{1-2\alpha}{2}\}}$ uniformly over a dense family of input spaces, achieving $O(N^{-\gamma})$ accuracy under suitable source conditions.
• Consistency with Physics-Informed Gaussian Processes: The closed-form solution coincides, under Gaussian kernel choice and regularization scaling, with the posterior mean in physics-informed GP regression, but achieves frequentist uniform norm convergence rather than Bayesian uncertainty quantification.

Numerical Experiments

Extensive computational results validate the approach's accuracy, efficiency, and robustness:

• Darcy Flow Problem: Achieves relative $L^2$ errors on the order $T_{\Lambda,N}$0 across smooth and heterogeneous permeability fields, with computational cost consistently below 2 seconds. The solution is robust to discontinuous source/boundary data.
• Helmholtz Equation (Comparison with Green Operator Learning): The proposed kernel operator method substantially outperforms the Green operator approach of [66], particularly in high-frequency regimes ($T_{\Lambda,N}$1), maintaining $T_{\Lambda,N}$2 error at $T_{\Lambda,N}$3 versus the competitor's failure with errors near 1. Computational savings are dramatic, with kernel methods requiring seconds versus hours for the Green operator learning due to avoidance of explicit paired data training.
• Supplementary Benchmarks: Additional tests on the Schrödinger and heat equations, including time-dependent and non-oscillatory regimes, consistently exhibit low approximation errors, rapid solution times, and scalability across varying regularity and structure in the input data.

Implications and Future Directions

This kernel-based operator learning paradigm theoretically and practically transcends traditional kernel methods and neural operators by:

• Enabling operator-level generalization independent of individual PDE instance training,
• Supporting efficient online updates and scalability via low-rank representations,
• Providing a rigorous statistical characterization with strong guarantees on convergence and robustness across source spaces.

From an applied standpoint, the method is immediately relevant for parametric PDE simulation, optimization, and uncertainty quantification in engineering—a domain where repeated operator evaluation is crucial. Potential extensions include structure-preserving kernel constructions for nonlinear PDEs, operator learning on manifolds, and integration of randomized low-rank techniques for further scalability.

Theoretically, the operator-theoretic viewpoint opens avenues for deeper spectral analysis of solution operators, adaptive regularization strategies, and connections with Bayesian approaches in uncertainty quantification.

Conclusion

The paper establishes a principled, scalable, and statistically robust kernel framework for learning PDE solution operators, demonstrating an explicit and efficient mechanism for operator generalization. Its performance advantage, especially in regimes demanding extrapolation and high fidelity, positions it as a compelling bridge between classical kernel numerics and contemporary operator learning, with both immediate and far-reaching implications for computational science and engineering (2605.09643).