Researchers Yan Ma and Wei Li have established a comprehensive benchmarking and theoretical framework demonstrating that the choice of numerical discretization strategy fundamentally determines the accuracy and stability of Physics-Informed Neural Networks when solving Fredholm integral equations of the second kind. Integral equations play a central mathematical role in describing physical processes across scientific and engineering disciplines, ranging from quantum field theory to computed tomography image reconstruction. While physics-informed neural architectures have emerged as flexible tools capable of integrating governing equations directly into neural learning objectives, existing implementations have predominantly relied on ad-hoc trial and error when discretizing integral operators. The new investigation demonstrates that discretization choice is not a secondary implementation detail but the dominant governing factor dictating numerical convergence, outperforming architecture tuning and offering substantial performance advantages over classical numerical solvers for challenging singular kernels.
Fredholm integral equations of the second kind define an unknown function through its relationship with an external driving term and an integral operator controlled by a bivariate kernel function. The mathematical structure of the kernel dictates the analytical properties of the solution, spanning infinitely smooth exponential interactions, weakly singular forms with algebraic branch-type infinities, and regularized kernels derived from inverse imaging transforms. To solve these equations within a neural framework, the continuous integral must be converted into discrete approximations suitable for evaluating residual loss functions at designated collocation points. The researchers systematically analyzed three foundational discretization strategies: the discrete coordinate method, which maintains continuous solution representations across discretized subintervals; the endpoint method, which evaluates functions at boundary nodes corresponding to composite trapezoidal integration; and the midpoint method, which computes kernel averages and evaluates functions at subinterval centers.
Theoretical analysis rooted in the functional framework of Lebesgue L1 spaces provided the mathematical underpinning for the comparative study. Formulating the problem in L1 space naturally accommodates discontinuous solutions and weakly singular kernels where classical square-integrable formulations encounter theoretical barriers. Under this framework, the integral operator is established as compact under standard operator assumptions, satisfying the Fredholm alternative theorem regarding existence and uniqueness. By analyzing approximation errors via Taylor expansions and Neumann series bounds, the researchers proved rigorous convergence theorems for each method. The discrete coordinate method was shown to achieve second-order convergence with an error bound scaling as O(h squared) for sufficiently smooth kernels, where h denotes the uniform mesh size. For weakly singular kernels characterized by power-law singularities of order alpha between zero and one, the endpoint method was proven to yield an optimal convergence rate scaling as O(h to the power of one minus alpha).
To isolate the precise numerical effects of discretization from confounding network variables, the investigators deployed a unified neural network architecture across all comparative experiments. The model utilizes a fully connected feedforward neural network comprising five hidden layers with 512 neurons per layer, yielding approximately 1.05 million trainable parameters. Hyperbolic tangent activation functions were selected specifically for their infinite smoothness, which ensures that network outputs and their corresponding integrals remain differentiable to arbitrary order, alongside a bounded range that provides natural regularization against gradient explosion during optimization. Network weights were initialized using Xavier normal initialization. Model training utilized the AdamW optimizer with an initial learning rate of 4 times 10 to the negative fourth power, a weight decay factor of 10 to the negative fifth power, and a cosine annealing schedule terminating at 10 to the negative sixth power, with gradient clipping enforced at a maximum norm of 1.0 to stabilize optimization across singular test cases.
The experimental protocol conducted on dedicated graphics processing hardware demonstrated that kernel regularity dictates the optimal choice of discretization technique. For smooth exponential kernels representing rapidly decaying interactions common in quantum mechanics and statistical physics, the discrete coordinate method achieved superior accuracy across all evaluated metrics. Preserving the analytical integral structure allowed the discrete coordinate approach to achieve a 35 percent reduction in physical residual compared to the endpoint method, recording a residual of 2.514 times 10 to the negative third power versus 3.287 times 10 to the negative third power on identical uniform discretizations of 50 intervals. Convergence tracking revealed that the discrete coordinate formulation reached its asymptotic error floor within 200 training epochs, maintaining higher-order structural fidelity without requiring excessively dense integration grids.
When evaluated on weakly singular kernels where traditional numerical quadrature rules typically break down, the comparative behavior shifted dramatically. Examining power-law singularities of the form where the kernel behaves inversely proportional to the square root of the coordinate separation, the endpoint method proved substantially superior. By explicitly evaluating singular intervals through composite trapezoidal formulations, the endpoint strategy achieved a 40 percent improvement in stability metric, scoring 0.89 compared to 0.74 for the discrete coordinate method. Regional error evaluations confirmed that this stability advantage manifested uniformly across the computational domain, maintaining a maximum pointwise error of 1.086 times 10 to the negative second power compared to 1.221 times 10 to the negative second power for discrete coordinates, with the most pronounced improvements concentrated directly along the singular diagonal.
The benchmark also incorporated rigorous comparisons against established classical numerical solvers, including the Nyström method implementing Gauss-Legendre quadrature, the Galerkin method employing a Legendre polynomial basis, and classical collocation based on trapezoidal quadrature. For well-behaved smooth kernels, classical solvers achieved superior pointwise precision because they compute exact matrix inversions directly without non-convex optimization overhead. However, on weakly singular test problems with a singularity parameter of alpha equal to 0.3, the physics-informed neural network midpoint method achieved the lowest absolute error among all evaluated techniques. Recording an error of 5.04 times 10 to the negative third power, the neural midpoint approach outperformed classical Galerkin methods by a factor of 2.6, classical collocation by 2.7, and Nyström solvers by 3.5, highlighting that neural formulations combined with structure-aware discretization avoid the severe numerical instabilities that hinder classical quadratures on non-smooth operators.
Empirical mesh refinement experiments systematically validated the theoretical convergence rates predicted by the mathematical proofs. Refining uniform meshes across discrete levels from 20 up to 320 nodes within the collocation framework produced error reduction ratios that mirrored theoretical exponents. For the smooth test kernel, successive mesh refinements yielded measured error reduction ratios of 4.18, 4.01, and 3.71, closely aligning with the theoretical value of 4.0 dictated by second-order O(h squared) convergence. For the weakly singular problem with alpha equal to 0.3, refinement steps produced measured ratios of 1.64, 1.63, 1.63, and 1.62, exhibiting near-perfect correspondence with the theoretical prediction of 2 to the power of 0.7, approximately 1.62, confirming that the analytical bounds accurately govern practical algorithm behavior.
Practical engineering efficacy was further evaluated on an inverse computed tomography reconstruction problem formulated around a regularized Abel transform with a smoothing parameter of 0.05. In this imaging application, the midpoint method delivered optimal reconstruction quality, achieving an accuracy enhancement of 25 percent and reaching a peak signal-to-noise ratio of 29.1 decibels. Complementary network ablation studies varying hidden layer depth from 3 to 7 layers and widths from 256 to 1024 neurons demonstrated that altering network capacity produced minimal accuracy variations compared to the substantial performance shifts caused by switching discretization strategies. Increasing network depth to 7 layers actually degraded accuracy due to optimization difficulties, confirming that mathematical discretization represents the primary governing factor in solving Fredholm integral equations via scientific machine learning.
Computational resource requirements introduce distinct trade-offs that practitioners must balance against desired accuracy. While the discrete coordinate method achieves optimal second-order convergence on smooth kernels, it requires storing intermediate interval integral caches, scaling memory consumption to O(nN) and computational operations to O(n squared N) for n discretization intervals and N collocation points. In contrast, point-evaluation schemes like endpoint and midpoint methods operate with streamlined memory footprints of O(n plus N) and operation counts of O(nN). Although the current empirical study focused primarily on one-dimensional test problems to maintain analytical tractability, the researchers noted that tensor-product extensions allow these kernel-adaptive principles to translate directly to higher-dimensional geometries and complex systems, providing a principled foundation to replace trial-and-error discretization in scientific computing.
Subject of Research: Kernel-adaptive discretization strategies for physics-informed neural networks solving Fredholm integral equations
Article Title: Kernel-Adaptive Discretization Strategies for Physics-Informed Neural Networks: A Comprehensive Framework for Optimal Solution of Fredholm Integral Equations
Article References: Ma, Y., & Li, W. (2026). Kernel-Adaptive Discretization Strategies for Physics-Informed Neural Networks: A Comprehensive Framework for Optimal Solution of Fredholm Integral Equations. Journal of Nonlinear Mathematical Physics, 33(1), Article 92. https://doi.org/10.1007/s44198-026-00450-5
Image Credits: AI Generated
DOI: 10.1007/s44198-026-00450-5
Keywords: Physics-Informed Neural Networks, Fredholm Integral Equations, Kernel-Adaptive Discretization, Discrete Coordinate Method, Endpoint Method, Midpoint Method, Weakly Singular Kernels, Smooth Kernels, Regularized Kernels, Scientific Machine Learning, Numerical Quadrature, Classical Solvers, Abel Transform, Computed Tomography Reconstruction, Computational Mathematics, Error Decomposition
Cite Scienmag News
Cassandra Pierce. (November 21, 2026). Kernel-Adaptive Discretization Improves Neural Network Solutions for Fredholm Integral Equations. Scienmag. https://scienmag.com/kernel-adaptive-discretization-improves-neural-network-solutions-for-fredholm-integral-equations/
Cassandra Pierce. "Kernel-Adaptive Discretization Improves Neural Network Solutions for Fredholm Integral Equations." Scienmag, 21 November 2026, https://scienmag.com/kernel-adaptive-discretization-improves-neural-network-solutions-for-fredholm-integral-equations/. Accessed 21 November 2026.
Cassandra Pierce. "Kernel-Adaptive Discretization Improves Neural Network Solutions for Fredholm Integral Equations." Scienmag. November 21, 2026. https://scienmag.com/kernel-adaptive-discretization-improves-neural-network-solutions-for-fredholm-integral-equations/








