WorksComputer Science
CFKS: Continuum-Feature Kernel PDE Solvers
Abstract
Neural-network-based PDE solvers require architectural choices, particularly the number of layers and neurons per layer. Fixing a single Gaussian radial basis function (RBF) feature layer and taking the number of features to the continuum limit removes the need to prescribe a finite feature count or select individual feature centers, thereby reducing representation arbitrariness. Motivated by this principle, we introduce Continuum-Feature Kernel Solver (CFKS), a deterministic framework for PDEs obtained by analytically marginalizing a bounded continuum of Gaussian RBF features. CFKS works directly with the resulting continuum-feature mean kernel and its operator-applied counterparts, parameterized by prescribed Gaussian widths and bounded center distributions. For linear PDEs, CFKS leads to symmetric operator-kernel collocation systems with null-space enforcement of physical constraints. Nonlinear PDEs are addressed through linearized initialization and nonlinear least-squares residual minimization using analytical kernel derivatives and Jacobians. Residual-guided collocation enrichment and sequential space–time-slab marching provide additional resolution strategies, while prescribed multiscale kernel refinement helps resolve localized nonlinear structures. Finite-feature experiments confirm convergence of empirical value and operator Gram matrices toward their continuum counterparts. Numerical studies include oscillatory and multiscale ODEs, irregular-domain Poisson and biharmonic problems, singular perturbations, transient transport equations, and viscous Burgers' equation. The results establish CFKS as a direct computational connection between finite Gaussian-RBF representations and continuum-feature kernel PDE solvers. Codes: https://github.com/vikas-dwivedi-2022/CFKS_Codes
Introduction
Physics-informed neural networks (PINNs) are increasingly used for forward and inverse partial differential equation (PDE) problems [1]. However, their practical performance depends on architectural choices, particularly the number of layers and neurons per layer, as well as optimization and collocation-point placement. A natural way to reduce this architectural arbitrariness is to consider a single-layer Gaussian radial basis function (RBF) representation and take the number of features to the continuum limit. When the corresponding feature expectation is available analytically, the resulting kernel eliminates the need to prescribe a finite feature count or select individual feature centers. The representation nevertheless remains parameterized by prescribed Gaussian widths and center distributions, which influence approximation accuracy and numerical conditioning. This motivates the development of meshfree PDE solvers that work directly with continuum-feature kernels rather than finite feature realizations.
A particularly relevant line of development begins with radial-basis-function (RBF) collocation. Kansa introduced meshfree PDE solvers based on finite RBF expansions with explicitly prescribed centers [2, 3]. Subsequent work provided approximation-theoretic foundations and clarified limitations of unsymmetric collocation systems [4, 5]. Symmetric and Hermite formulations subsequently incorporated differential and boundary operators directly into kernel collocation by applying the corresponding operators to both arguments of a positive-definite kernel, thereby constructing symmetric operator-based collocation matrices [6].
In parallel, Gaussian-process and neural-network theory established the correspondence between feature representations and kernels. Infinite-width neural models may converge to Gaussian processes [7], while Gaussian-process theory relates continuous basis-function representations to covariance kernels [8]. In such representations, a deterministic kernel can be constructed by averaging products of basis functions over a prescribed distribution of feature parameters. We refer to the resulting kernel as the continuum-feature mean kernel. Neither infinite-feature limits nor continuum Gaussian-basis representations are new in themselves. The question addressed in this work is how to evaluate the bounded Gaussian-RBF mean kernel and its differential-operator counterparts analytically and employ them directly in PDE solvers with prescribed physical constraints.
Random-feature methods exploit the converse direction by approximating this mean kernel using a finite sampled feature map [9]. This idea has been transferred to PDEs through random-feature methods that retain a finite randomized trial space and solve only for output coefficients [10]. More recent work has considered localized, high-order, and operator-aware variants, including residual-Christoffel sampling for finite operator-applied feature spaces [11] and analytic partition-of-unity constructions for high-order PDEs [12]. Kernel and Gaussian-process methods have likewise been applied directly to differential operators, probabilistic collocation, and structurally constrained boundary-value problems [13, 14, 15].
The present work connects these developments through Continuum-Feature Kernel Solver (CFKS), based on the exact bounded mean kernel induced by a Gaussian RBF feature family. Unlike finite RBF representations, which require a prescribed number of features and their individual center locations, CFKS replaces the finite feature realization with analytical averaging over a prescribed distribution of centers. This produces a deterministic continuum-feature mean kernel, eliminating finite-feature sampling variability while retaining the Gaussian widths and center distributions as prescribed parameters.
An important property of this construction is that differential operators can be applied directly to the mean kernel. The resulting operator-applied kernels are equivalent to analytically averaging the products of the corresponding operator-applied Gaussian features. Consequently, the kernel matrices required for PDE collocation can be constructed without explicitly generating finite RBF features. For Gaussian RBFs with centers uniformly distributed over bounded intervals or boxes, the mean kernel admits a closed-form expression involving a squared-exponential factor and error-function terms that account for the bounded center support [8].
For linear PDEs, CFKS uses the continuum-feature mean kernel and its operator-applied counterparts to construct symmetric strong-form collocation systems. Boundary and initial conditions can be enforced through a null-space parameterization, separating the satisfaction of physical constraints from the interior PDE approximation. For nonlinear PDEs, analytical kernel derivatives enable direct evaluation of the nonlinear residual and its Jacobian. Linearized solves provide initialization for nonlinear least-squares residual minimization, while analytical trial-function parameterizations can enforce the prescribed physical conditions. Problems exhibiting localized structures can be addressed through multiscale continuum-feature kernels and residual-guided collocation enrichment. These extensions are demonstrated using viscous Burgers’ equation with a two-stage predictor–corrector strategy.
Research gap Despite substantial developments in RBF collocation, kernel methods, Gaussian-process formulations, and random-feature PDE solvers, two methodological gaps motivate the present work:
Finite RBF and random-feature PDE methods generally require a prescribed number of features and an explicit realization of their centers, whereas conventional kernel collocation typically starts from a prescribed kernel. Although continuum-feature limits are well established, the direct construction of PDE solvers from analytically marginalized bounded Gaussian-RBF feature distributions, including the corresponding differential-operator kernels, remains comparatively underexplored. Such a formulation removes the finite feature count and individual center selection while retaining prescribed Gaussian widths and center distributions.
The enforcement of boundary and initial conditions remains an important consideration in meshfree PDE solvers. Physical constraints are frequently incorporated through penalty terms, weighted least-squares formulations, or jointly assembled constraint systems. Although null-space methods for enforcing linear constraints are well established [16, 17, 18], their integration with analytically derived continuum-feature
operator kernels, together with analytical constraint embeddings for nonlinear problems, provides a unified approach for separating the enforcement of physical conditions from the approximation of the governing PDE.
Contributions The main contributions of this work are:
We develop CFKS by analytically replacing a finite Gaussian-RBF feature representation with its exact bounded continuum-feature mean kernel. This removes the need to prescribe a finite feature count or select individual feature centers, while retaining Gaussian widths and center distributions as prescribed parameters.
We derive the corresponding differential-operator kernels and employ them in symmetric strong-form collocation for linear PDEs. Boundary and initial conditions are enforced through a null-space parameterization, allowing the governing equations to be solved within a constraint-compatible subspace.
We extend CFKS to nonlinear PDEs through linearized initialization and nonlinear least-squares residual minimization, using analytical kernel derivatives and Jacobians. Analytical constraint embeddings and multiscale continuum-feature kernels further accommodate prescribed physical conditions and localized solution structures.
We integrate optional residual-adaptive collocation and sequential space–time-slab marching into the CFKS framework. Residual-guided enrichment concentrates computational resolution in regions of under-resolution, while prescribed multiscale kernel refinement provides additional flexibility for resolving localized structures.
Residual-based adaptive sampling is already well established in meshfree and physics-informed methods [19, 20]. In CFKS, it is incorporated as an optional refinement strategy in which independently evaluated PDE residuals guide the allocation of additional collocation points. When the mean kernel is fixed, this refinement changes the collocation and representer sites without altering the underlying continuum-feature kernel or prescribed physical conditions. For nonlinear problems, collocation enrichment can additionally be combined with prescribed changes in kernel scales. Sequential space–time-slab marching extends this strategy to time-dependent problems, consistent with earlier time-marching and causality-aware formulations [21, 22].
Organization of the paper The remainder of this paper is organized as follows. § 2 presents the mathematical formulation of CFKS, including the continuum-feature mean kernel, operator-based collocation, constraint enforcement, and extensions to nonlinear PDEs, residual-adaptive collocation, and sequential space–time slabs. § 3 introduces the numerical benchmark problems. § 4 presents and discusses the numerical results, including finite-feature convergence and conditioning studies. § 5 summarizes the main findings and outlines directions for future research. The appendices provide the connection between finite-feature models and the CFKS representer formulation, together with the analytical derivation of the bounded Gaussian continuum-feature kernel.
Mathematical Formulation
We first develop CFKS for linear differential problems and subsequently describe its extension to nonlinear PDEs. Let denote the computational domain and the vector of independent variables. For time-dependent problems, , where represents the spatial coordinates and denotes time. We consider the linear differential problem
where is a prescribed linear differential operator, is the unknown solution, and is a known source function. The problem is supplemented by boundary conditions and, for time-dependent equations, appropriate initial conditions.
To formulate the numerical problem, the governing PDE and its physical conditions are expressed through a collection of linear constraint functionals,
where is the total number of discrete constraints, denotes the linear functional associated with constraint , and is its prescribed value. For an interior PDE constraint imposed at a collocation point , the functional applies the differential operator to the solution at that point, with . Boundary and initial conditions are represented by the corresponding evaluation or differential functionals at the selected physical constraint locations.
From finite RBF features to the continuum-feature mean kernel
We begin with a finite representation of the unknown solution using Gaussian radial basis function (RBF) features. Let denote a point in the computational domain, where is the number of independent variables. For time-dependent problems, these variables include both spatial coordinates and time.
The th Gaussian RBF feature is defined as
where denotes the center of the th feature, is its center coordinate in the th direction, and is the prescribed Gaussian width associated with that direction. The index runs over the independent variables. Equal widths in all directions give the isotropic Gaussian RBF, whereas different widths allow coordinate-dependent scaling. In space–time problems, separate spatial and temporal widths can therefore be prescribed.
Using such features, where is the total number of Gaussian RBF features, the finite-feature approximation is expressed as
where denotes the approximate solution, indexes the individual features, and is the output coefficient associated with the th feature. The feature centers and Gaussian widths are fixed when determining the coefficients from the governing constraints. The normalization factor ensures that the associated feature Gram matrix has a nontrivial finite limit as the number of features increases.
Recall that the governing PDE and physical conditions are represented by linear constraint functionals, denoted by , with . Applying these functionals to the finite-feature representation gives the constraint matrix , whose entries are
where is the entry associated with constraint and feature . Thus, is an matrix, and denotes the result of applying the th linear constraint functional to the th Gaussian RBF feature.
The corresponding constraint Gram matrix is obtained by multiplying by its transpose:
where denotes the transpose of , and index pairs of constraint functionals. The matrix is symmetric and positive semidefinite, with each entry representing the inner product of two constraint-applied finite-feature vectors.
To obtain the continuum-feature formulation, suppose that the Gaussian RBF centers are independently sampled from a prescribed probability distribution with density , where represents a generic feature-center vector. The distribution specifies the region and relative weighting of admissible feature centers. In the present work, bounded uniform center distributions are used.
Rather than retaining an explicit finite realization of these centers, we consider the analytical mean of the products of Gaussian RBF features. This defines the continuum-feature mean kernel
where denotes the mean kernel associated with the center distribution , and and are its first and second arguments, respectively. The notation denotes expectation with respect to the prescribed probability distribution of the feature-center vector . The function is the Gaussian RBF defined in eq. 3, with the generic center replacing the individual center .
For independently sampled feature centers, the law of large numbers implies that, as the number of features tends to infinity, each entry of the empirical constraint Gram matrix converges to its corresponding population value, provided the relevant feature moments are finite. For the bounded Gaussian-feature construction considered here, these conditions are satisfied. Consequently,
where the superscripts on the constraint functionals identify the kernel argument on which each functional acts. Specifically, acts on the first argument of the kernel, while acts on the second. Each functional includes the appropriate differential or evaluation operation at the location associated with its constraint.
Equation (8) establishes that the continuum-feature constraint Gram matrix is the analytical expectation of the corresponding finite-feature Gram matrix. Thus, CFKS replaces the finite feature count and individual center realization with a deterministic kernel defined by the prescribed center distribution and Gaussian widths.
For Gaussian RBF features with centers uniformly distributed over bounded intervals, the expectation defining the continuum-feature mean kernel can be evaluated analytically. In one dimension, the resulting expression consists of a Gaussian factor multiplied by error-function terms that account for the finite support of the center distribution. When the center distribution is a product of independent uniform distributions over a Cartesian box, the multidimensional mean kernel is obtained as the product of the corresponding one-dimensional kernels. The complete derivation and closed-form expressions are provided in Appendix § 7.
An important property of the continuum-feature mean kernel is that its derivatives can be obtained directly from the analytical kernel expression. To describe derivatives of arbitrary order, let and denote two multi-indices, where and are nonnegative integers specifying the orders of partial differentiation with respect to the th coordinates of and , respectively. The operators and denote the corresponding mixed partial derivatives with respect to the first and second kernel arguments.
For the smooth Gaussian features and bounded center distributions considered here, differentiation can be interchanged with expectation. Therefore,
which shows that the mixed derivatives of the mean kernel are equal to the analytical averages of the corresponding products of differentiated Gaussian RBF features.
In particular, for the prescribed linear differential operator appearing in the governing PDE,
where and indicate application of the differential operator to the first and second kernel arguments, respectively. On the right-hand side, denotes the operator applied to a Gaussian RBF feature and evaluated at , with the analogous interpretation for the second argument.
Equation (10) establishes that the operator-applied mean kernel is exactly the population Gram kernel associated with the corresponding operator-applied Gaussian RBF features. Hence, the differential-operator kernels required by CFKS follow directly from the continuum-feature construction rather than being introduced independently.
Linear CFKS formulation
We now construct the CFKS approximation for the linear differential problem introduced in eq. (1). Recall that , with , denotes the linear functional associated with the th governing PDE, boundary, or initial constraint, and is the total number of discrete constraints. Each functional includes the appropriate differential or evaluation operation at its prescribed constraint location.
Let denote the continuum-feature mean kernel defined in eq. (7). The CFKS approximation is expressed in the kernel representer form
where is the kernel approximation to the unknown solution, is an arbitrary evaluation point, is the coefficient associated with the th constraint, and is the second argument of the kernel. The superscript on indicates that the th constraint functional acts on the second kernel argument, including evaluation at the location associated with that constraint. When the functional contains derivatives, these are applied before evaluating at the corresponding constraint location.
The representer formulation follows from the connection between finite-feature models and their continuum-kernel counterparts, as detailed in Appendix § 6.
Applying each constraint functional to eq. (11) produces the linear system associated with the discrete PDE and physical conditions. Its coefficient matrix is the operator Gram matrix , with entries
where and act on the first and second kernel arguments, respectively. Both functionals include evaluation at their prescribed constraint locations. Thus, represents the interaction between the th and th constraint functionals through the continuum-feature mean kernel.
The matrix has dimensions and is symmetric and positive semidefinite. As established in eq. 8, its entries are the analytical population limits of the corresponding finite-feature constraint Gram-matrix entries.
Let denote the vector of representer coefficients and the vector of prescribed constraint values. A regularized collocation formulation is then given by
where is the identity matrix and is a prescribed regularization parameter. When , the system corresponds to direct constraint interpolation, provided a solution exists. A positive regularization parameter improves the conditioning of the coefficient system but generally relaxes exact satisfaction of the discrete constraints. The separate enforcement of linear physical conditions through a null-space parameterization is described in the subsequent formulation.
Once the coefficient vector has been determined, the approximate solution can be evaluated at any point using
where is an -component column vector containing the kernel interactions between the evaluation point and the selected constraint functionals. Its th component is obtained by applying the th functional to the second argument of the continuum-feature mean kernel. The superscript denotes vector transposition.
Equations 7–14 define the kernel formulation without retaining the finite number of Gaussian RBF features or their individual center realizations. Because the bounded continuum-feature mean kernel is evaluated analytically, no numerical quadrature over the feature-center distribution is required. The resulting PDE approximation is determined by the prescribed kernel parameters, the selected constraint locations, and the numerical treatment of the resulting collocation system.
Constraint-compatible enforcement of physical conditions
When boundary or initial data are to be satisfied independently of the PDE least-squares fit, we separate the physical constraints from the interior PDE equations. Let
denote, respectively, the interior PDE equations and the boundary or initial constraints obtained from the CFKS representation eq. 11.
A particular coefficient vector satisfying the physical constraints is chosen as
where denotes the Moore–Penrose pseudoinverse. Let the columns of span the numerical null space of ,
Every constraint-compatible solution can then be parameterized as
Substitution into the PDE equations gives the reduced problem
With Tikhonov regularization,
This construction separates satisfaction of the linear physical constraints from fitting of the PDE residual and avoids introducing an arbitrary relative penalty between boundary and interior equations. In practice, the reduced system is solved using rank-revealing SVD-based linear algebra.
Nonlinear CFKS formulation
We now extend the continuum-feature kernel formulation to nonlinear differential problems. Consider
where is a prescribed nonlinear differential operator, is the unknown solution, is a known source function, and is the computational domain. The operator may depend nonlinearly on the solution, its derivatives, or both. Appropriate boundary and, when applicable, initial conditions complete the problem.
Unlike the linear formulation, a nonlinear differential operator does not generally produce a fixed operator Gram matrix. Instead, CFKS constructs a deterministic kernel trial representation and determines its coefficients by minimizing the nonlinear PDE residual.
Let be the continuum-feature mean kernel defined in eq. (7). We consider the trial representation
where denotes the approximate solution, is the unknown coefficient vector, is the number of kernel representer sites, and is the location of the th representer site. The prescribed functions and define an additive component and a multiplicative factor, respectively.
When a suitable analytical constraint embedding is available, these functions can be selected so that the trial solution satisfies the prescribed boundary or initial conditions independently of the coefficient values. An unconstrained representation is recovered by setting and .
The trial representation is linear in , although the governing PDE residual is generally nonlinear in these coefficients. Importantly, determines the dimension of the discrete kernel representation and does not represent the number of Gaussian RBF features. Each kernel section is already obtained by analytically averaging a continuum of Gaussian RBF features over the prescribed center distribution.
Let denote the interior PDE collocation locations, where is the number of discrete PDE constraints. The residual at the th collocation point is defined as
where the nonlinear operator acts on the trial solution before evaluation at .
Collecting the residuals into the vector , the kernel coefficients are determined through
where denotes the optimized coefficient vector and is the Euclidean norm. Thus, the nonlinear PDE is enforced through least-squares minimization of its discrete residual.
The analytical continuum-feature kernel also permits direct construction of the residual Jacobian. Define the coefficient-independent trial functions
Since the derivative of the trial solution with respect to is , the Jacobian entries are
provided the nonlinear operator is Fréchet differentiable. Here, denotes the linearization of about the current trial solution, acting on the function . The Jacobian has dimensions .
Because the mean kernel and its derivatives are available analytically, the residual and Jacobian can be evaluated without generating finite random-feature realizations. The nonlinear optimization therefore determines only the coefficients of the prescribed continuum-feature kernel representation.
Linearized initialization. The nonlinear least-squares problem generally requires an initial coefficient vector. A suitable initialization can be obtained by linearizing the governing equation about an available approximation.
Let denote the current approximation at iteration , and let be a correction. A first-order expansion gives
where is the Fréchet derivative of the nonlinear operator evaluated at . The corresponding linearized equation is
For a fixed approximation , this is a linear differential problem for the correction . Its discretization can use the analytical kernel derivatives developed for linear CFKS.
Alternatively, a Picard-type approximation may be used when the nonlinear operator admits an appropriate coefficient-freezing formulation. The resulting linearized solution supplies initial coefficients for the nonlinear least-squares problem eq. 24.
This initialization procedure does not replace the nonlinear coefficient solve. Rather, it provides a starting approximation for minimizing the full nonlinear PDE residual.
Multiscale continuum-feature kernels For problems with localized gradients or multiple characteristic scales, the trial representation can combine continuum-feature mean kernels with different prescribed Gaussian widths and center distributions.
We consider an additive global–local kernel of the form
where is the multiscale kernel at refinement stage , and is a prescribed positive mixing weight. The subscripts and identify the global and local components, respectively.
The global kernel is constructed from a bounded distribution of Gaussian centers covering the spatial domain, whereas the local kernel uses a center distribution concentrated around a prescribed region of interest. Each component is an analytically marginalized Gaussian-RBF mean kernel of the form defined in eq. 7.
To balance their contributions, the components are normalized according to
where identifies the kernel component and is a prescribed reference point at which the diagonal kernel value is positive.
Positive normalization factors preserve positive semidefiniteness, and the positive weighted sum in eq. 29 remains a symmetric positive-semidefinite kernel. Its derivatives are obtained by differentiating the analytical component kernels.
The center distributions, Gaussian widths, normalization point, and mixing weight remain prescribed parameters. The multiscale construction therefore increases representation flexibility without introducing finite Gaussian-feature center sampling.
Illustration using viscous Burgers’ equation As a nonlinear example, consider the one-dimensional viscous Burgers equation
where denotes the spatial coordinate, is time, is the final simulation time, and is the viscosity. The subscripts , , and denote the corresponding first time, first spatial, and second spatial derivatives.
For the homogeneous Dirichlet boundary conditions and prescribed initial profile used in this benchmark, analytical constraint embedding is applied on successive space–time slabs.
Let index the time slabs, with denoting the local time measured from the beginning of the current slab. At refinement stage , the trial solution is
where is the incoming solution profile and is an unknown correction field.
The factors and make the correction vanish at the beginning of each slab and at the spatial boundaries, respectively. Hence, the trial solution preserves the incoming profile and satisfies the homogeneous boundary conditions, provided the incoming profile is boundary-compatible.
The correction field is represented by
where , is the number of PDE collocation points on slab at stage , denotes the corresponding collocation location, and is its kernel coefficient. Thus, the PDE collocation locations also serve as kernel representer sites.
For a fixed slab and stage, let collect the unknown coefficients, with the slab and stage indices suppressed for simplicity. Analytical differentiation of the embedded trial solution gives the affine relations
where the vectors on the left contain the solution and its derivatives at the PDE collocation points. The vectors , , and contain the incoming profile and its spatial derivatives at those points.
The matrices , , , and represent the corresponding contributions of the kernel correction, including derivatives of the analytical embedding factor. The incoming profile has no local-time dependence and therefore contributes nothing to the time derivative.
The discrete nonlinear residual is
where denotes componentwise multiplication. Differentiating this residual with respect to the kernel coefficients yields
where denotes the diagonal matrix formed from a vector.
The nonlinear coefficients are obtained by minimizing the residual in eq. (35) using the analytical Jacobian in eq. (36). Linearized initialization, multiscale kernel refinement, and residual-guided collocation are used in the numerical implementation, with the specific settings described in the benchmark setup.
Residual-adaptive collocation
The continuum-feature kernel formulation permits either fixed or adaptive PDE collocation. A fixed deterministic point distribution is sufficient when it adequately resolves the solution. For problems with localized under-resolution, residual-adaptive collocation can concentrate additional PDE constraints in regions where the current approximation exhibits larger governing-equation residuals.
This refinement mechanism is not required by CFKS itself. It provides an additional resolution strategy that can be combined with linear or nonlinear kernel formulations.
To describe both cases, define the PDE residual operator
where and are the prescribed linear and nonlinear differential operators, respectively, and is the corresponding source function.
Let denote an initial deterministic PDE collocation set, and let be the CFKS approximation obtained from this set. We refer to as the predictor solution.
The computational region is partitioned into blocks, denoted by , where is the block index. For each block, define an independent residual probe set . The probe points are distinct from the PDE collocation locations used to construct the predictor.
At a probe point , the predictor residual is
where measures the discrepancy between the governing differential equation and the predictor approximation.
The residual magnitude within each block is characterized by its root-mean-square and maximum values,
Here, is the number of probe points in block , measures the average residual magnitude, and measures the largest observed residual.
A normalized residual indicator is constructed by combining these measures:
where is the indicator for block , is a prescribed weighting parameter, and prevents division by zero. The index runs over all blocks in the region being refined. The two terms measure, respectively, the relative average residual and the relative peak residual.
The indicator determines the refinement level assigned to each block. For a three-level strategy, define
where denotes the refinement level, and and are prescribed thresholds satisfying .
The levels correspond to prescribed collocation budgets , , and . The number of PDE collocation points assigned to block , denoted by , is selected according to its refinement level . The total number of PDE collocation points after refinement is therefore
where is the total refined collocation budget.
Additional implementation rules may be introduced to avoid abrupt spatial changes in collocation density, such as extending high-refinement assignments to neighboring blocks. The particular thresholds, point budgets, and neighboring-block rules are selected for the numerical problem under study.
Once the blockwise budgets are determined, the refined PDE collocation locations are generated within the corresponding blocks. Deterministic low-discrepancy sequences provide one means of constructing these point sets without introducing random sampling.
For a prescribed continuum-feature mean kernel, residual-guided collocation refinement modifies the discrete PDE constraints but does not change the analytical kernel itself. If the kernel representer sites coincide with the collocation locations, as in the nonlinear implementation considered here, enrichment also changes the number of kernel coefficients.
Residual-guided collocation may additionally be combined with prescribed changes in Gaussian kernel widths or center distributions. In that case, the discrete collocation representation and the underlying kernel parameters are both modified.
In the viscous Burgers experiment, residual-adaptive collocation forms part of the second stage of a predictor–corrector procedure. Residual indicators computed from the completed predictor determine the corrector collocation distribution. The corrector also employs narrower prescribed kernel widths, so the resulting accuracy improvement cannot be attributed to collocation adaptation alone.
The numerical values used for the residual weights, refinement thresholds, probe distributions, and collocation budgets are specified with the corresponding test problem.
Residual-adaptive collocation settings The two-stage Burgers computation uses a deterministic predictor followed by a residual-guided corrector. The spatial interval is divided into 20 equal blocks. For each time slab, the predictor uses 10 PDE collocation points per block, giving 200 points per slab.
The predictor residual is evaluated independently on a tensor-product probe grid within each block, using 11 spatial locations and 7 local-time locations. Spatial probe positions extend from 5% to 95% of the block width, while the local-time positions range from to , where is the slab duration. Thus, each block contains 77 residual probes.
The residual indicator in eq. (40) uses , assigning weights 0.7 and 0.3 to the root-mean-square and maximum residual measures, respectively. The refinement thresholds in eq. (41) are prescribed as and .
To reduce abrupt changes in collocation density, high- and medium-refinement assignments are extended to immediately adjacent spatial blocks. The expansion does not wrap around the domain boundaries, and high-refinement assignments take precedence where the expanded regions overlap.
The corrector uses the blockwise collocation budgets
The residual normalization and refinement decisions are performed separately for each time slab. Consequently, the corrector collocation count varies with the residual distribution.
Collocation points in both stages are generated using a deterministic two-dimensional Halton sequence with bases 2 and 3. Within each spatial block, the first Halton points are mapped to the corresponding spatial and local-time intervals, where is the prescribed blockwise point budget.
The Stage 2 corrector combines this residual-guided collocation enrichment with narrower prescribed global and local Gaussian kernel widths. The kernel representer sites coincide with the PDE collocation locations in both stages.
Sequential space–time-slab formulation
For time-dependent PDEs, CFKS can be applied sequentially over consecutive space–time slabs rather than constructing a single kernel approximation over the entire temporal domain. This strategy reduces the size of the individual space–time problems and allows the solution to be advanced through a sequence of local temporal intervals.
Let denote the spatial domain and the temporal interval, where is the initial time and is the final time. The temporal interval is partitioned into consecutive slabs,
where is the slab index, denotes the total number of slabs, and the partition points satisfy
The duration of slab is .
Within each slab, the physical time coordinate is replaced by the local coordinate
where and correspond to the beginning and end of the current slab, respectively.
The CFKS approximation on slab is denoted by , where represents the spatial coordinates. The governing PDE is enforced at selected collocation points within the corresponding space–time region . The same continuum-feature kernel formulation developed in the preceding sections applies to each slab.
The local coordinate system permits the same analytical mean-kernel construction to be used across different slabs. When the slab durations, local collocation configurations, and prescribed kernel parameters are identical, the corresponding kernel matrices and analytical derivative evaluations may also be reused.
Sequential transfer of the solution The initial condition for each slab is obtained from the prescribed initial data or from the terminal solution of the preceding slab.
Let denote the prescribed initial profile at time . The sequential transfer condition is
where represents the terminal profile of the preceding slab.
Thus, each local space–time problem receives its initial data from the preceding time interval. The solution is advanced sequentially until the final time is reached.
The incoming profile may be imposed through discrete initial constraints or incorporated analytically into a constraint-compatible trial representation. In either case, the local approximation is constructed to match the prescribed incoming data according to the chosen constraint-enforcement procedure.
In numerical implementations, the terminal profile may be transferred directly or represented through an interpolation procedure. When interpolation is used, the incoming profile is an approximation of the preceding terminal solution, and the transfer accuracy becomes part of the overall temporal discretization error.
Within each slab, the CFKS representation remains global over the spatial domain. The temporal decomposition therefore does not introduce artificial spatial interfaces or require spatial domain decomposition.
Predictor–corrector slab marching Sequential slab marching can also be combined with residual-adaptive collocation. In this case, a predictor solution is first computed using an initial deterministic collocation distribution.
Let denote the predictor approximation on slab , where the superscript (0) identifies the predictor stage. The predictor satisfies the sequential transfer condition
The predictor is advanced through the entire temporal interval, with each slab initialized from the preceding predictor slab.
After completing the predictor march, the PDE residual is evaluated at independent probe locations, as described in Section 2.5. The resulting residual indicators identify regions requiring additional collocation resolution.
Let denote the refinement level assigned to block on slab , where is the spatial-block index. These refinement assignments determine the collocation distribution used for the subsequent corrector computation.
The corrector solution, denoted by , is then computed through a second sequential march, with
where the superscript (1) identifies the corrector stage.
The predictor and corrector marches are performed independently, each beginning from the prescribed initial condition. The predictor residual information determines the refined collocation distributions, but the predictor solution is not imposed as the initial profile of the corrector.
Instead, continuity between successive corrector slabs is maintained through the terminal profile of the preceding corrector slab, subject to the selected numerical transfer procedure.
The corrector may retain the prescribed mean kernel used by the predictor or employ different prescribed kernel parameters. In the latter case, residual-guided collocation refinement and kernel-scale refinement are combined within the same sequential solution strategy.
General characteristics Sequential space–time-slab marching is compatible with both linear and nonlinear CFKS formulations. The governing PDE is discretized independently within each slab, while the incoming solution profile supplies the initial data required for temporal advancement.
For linear problems, the local coefficients can be determined through the corresponding kernel collocation or constraint-compatible linear least-squares formulation. For nonlinear problems, the same slab decomposition can be combined with linearized initialization and nonlinear residual minimization.
The continuum-feature mean kernel and its analytical differential operators remain applicable within every slab. Temporal decomposition changes the organization of the discrete PDE solve without reintroducing finite Gaussian RBF feature realizations.
Residual-guided predictor–corrector marching is an additional refinement strategy rather than a requirement of the sequential formulation.
Test Cases
We consider a set of ODE and PDE benchmarks spanning oscillatory and multiscale one-dimensional problems, irregular two-dimensional geometries, a fourth-order equation, a singularly perturbed boundary layer, time-dependent transport, and a nonlinear viscous conservation law. The governing equations, computational domains, physical conditions, and reference solutions are summarized in Table 1. For the Poisson and biharmonic problems, the five-petal domain is centered at and its boundary is defined by for .
| Problem | Governing equation and domain | Conditions | Reference solution |
| First-order ODE | |||
| Second-order ODE | |||
| Multiscale ODE | |||
| Poisson | on the five-petal domain | on | |
| Biharmonic | on the five-petal domain | on | |
| Singular perturbation | |||
| Linear advection | |||
| Advection–diffusion | |||
| Viscous Burgers | Cole–Hopf reference solution |
Table 1. Summary of the numerical test problems.
Results and Discussion
ODE benchmarks
Figure 1 summarizes the three one-dimensional ODE benchmarks. In all cases, the continuum-feature kernel solution closely tracks the exact solution over the full interval. For the first-order oscillatory problem, the solution RMSE and maximum error are and , respectively. The second-order problem gives an RMSE of and a maximum error of , while the multiscale case yields and .

Figure 1. ODE benchmarks. Top: first-order oscillatory problem . Middle: second-order oscillatory problem . Bottom: multiscale problem with . In each row, the left panel compares the continuum-feature kernel and exact solutions, while the right panel shows the pointwise absolute error.
The corresponding residual RMS values are , , and for the first-order, second-order, and multiscale problems, respectively. In each case, the pointwise error is smallest at the constrained location and tends to increase toward the ends of the interval. The imposed value constraint is satisfied to machine precision in all three cases; for the second-order problem, the derivative condition is likewise recovered to machine precision.
Overall, these results verify the first- and second-derivative population kernels and show that the same continuum-feature representation can resolve both rapidly oscillating and multiscale solutions without introducing separate finite feature sets.
Table 2 reports the computational cost of the principal stages of the continuum-feature solve. For all three problems, the dominant cost is the assembly of the operator-kernel matrices, followed by the SVD of the reduced system. The null-space construction contributes only a small fraction of the total cost. The second-order problem is moderately more expensive because additional differentiated kernel blocks are required to represent both the second-order differential operator and the two physical constraints.
| Case | Kernel assembly (s) | Null-space (s) | Reduced SVD/solve (s) | Total solve (s) |
| First-order ODE | 1.518 | 0.022 | 0.603 | 2.155 |
| Second-order ODE | 2.427 | 0.036 | 0.823 | 3.296 |
| Multiscale ODE | 1.778 | 0.004 | 0.759 | 2.546 |
Table 2. Computational timing for the three ODE benchmarks.
Comparison with PINNs and FBPINNs
To contextualize the performance of CFKS, we compare it with standard physics-informed neural networks (PINNs) and finite-basis PINNs (FBPINNs). The three ODE benchmarks considered above coincide with those studied by Moseley et al. [23], allowing comparison under the same operator families and frequency content.
Standard PINNs exhibit pronounced spectral bias on these problems, particularly for the and multiscale cases. FBPINNs improve the high-frequency representation by decomposing the domain into overlapping subdomains with locally normalized subnetworks, but require the selection of the subdomain layout, overlap, network architecture, and training schedule [23]. For the first-order oscillatory problem, the reported FBPINN configuration uses 30 subdomains and 50,000 Adam iterations, whereas the second-order problem requires a specialized sequential “learning-outwards” schedule extending to approximately iterations [23].
CFKS does not employ gradient-based training or an explicit finite feature realization. The three problems are solved through deterministic population-kernel assembly followed by the null-space and reduced SVD solves. The measured total assembly-and-solve times are approximately 2.15, 3.30, and 2.55 s for the first-order, second-order, and multiscale problems, respectively. The corresponding solution RMSEs are , , and . Table 3 summarizes the comparison.
| Method | Training / solution cost | Architectural complexity | Accuracy on high-frequency and multiscale tests |
| PINN | Approximately – gradient steps; convergence deteriorates for high-frequency problems | Single neural network; performance depends on depth, width, activation choice, and may benefit from Fourier features | Large errors for relative to FBPINNs; affected by spectral bias [23] |
| FBPINN | Approximately – gradient steps depending on the problem | Typically 20–30 overlapping subdomains with multiple local networks; requires subdomain, overlap, and training-schedule choices | Accurately resolves the high-frequency and multiscale solutions; best reported first-order errors are of order [23] |
| CFKS | No backpropagation; deterministic kernel assembly and linear-algebra solve. Total solve times: 2.15–3.30 s | No neural architecture, finite hidden-layer width, or sampled center realization; prescribed kernel parameters and collocation constraints | Solution RMSEs of , , and for the first-order, second-order, and multiscale problems, respectively |
Table 3. Qualitative and quantitative comparison of PINN, FBPINN, and CFKS on the oscillatory and multiscale one-dimensional ODEs. PINN and FBPINN results are taken from (author?) [23].
The comparison highlights a different computational route to high-frequency resolution. FBPINNs address spectral bias by decomposing the domain and training multiple local neural networks, whereas CFKS works directly with the deterministic population kernel associated with the Gaussian feature family. Consequently, the three ODEs require no backpropagation, neural architecture selection, or finite-center sampling. The present timings should not be interpreted as a direct hardware-normalized speed comparison with the results of [23]; rather, they show that the population-kernel formulation attains comparable high-frequency resolution through a small number of dense linear-algebra operations.
Poisson and biharmonic problems on the five-petal domain
Figure 2 summarizes the CFKS results for the Poisson and biharmonic problems on the irregular five-petal domain. The top row corresponds to the second-order Poisson problem and the bottom row to the fourth-order biharmonic problem.

Figure 2. CFKS results on the five-petal domain. Top row: Poisson problem. Bottom row: biharmonic problem. In each row, the panels show the collocation geometry, CFKS solution, pointwise absolute error, and absolute PDE residual.
For the Poisson case, CFKS reproduces the manufactured solution over the entire nonconvex domain with an independent solution RMSE of and a maximum absolute error of . The maximum boundary error is , confirming that the Dirichlet constraint is maintained accurately by the null-space formulation. The independent PDE residual has RMS and maximum magnitude . As seen in the top row of Figure 2, the largest solution errors remain localized within the interior rather than being distributed uniformly over the domain.
The biharmonic problem provides a more demanding test because it requires fourth-order population-kernel derivatives together with simultaneous Dirichlet and normal-derivative boundary conditions. CFKS achieves an independent solution RMSE of and a maximum absolute error of . The maximum training boundary error is , while the dense Dirichlet-boundary error is . The biharmonic residual has RMS and maximum magnitude . Although the absolute residual is larger than in the Poisson case, this reflects the substantially larger scale introduced by repeated differentiation in the fourth-order operator; the solution itself remains accurate throughout the domain. The total CFKS solve time is 0.736 s, comprising approximately 0.663 s for kernel/operator assembly, 0.046 s for the null-space construction, and 0.020 s for the reduced SVD solve. The dominant computational cost is therefore the assembly of the high-order operator-kernel matrices rather than the constrained linear solve.
The two cases therefore show that the same CFKS construction extends from second- to fourth-order operators on irregular geometries without altering the underlying population-kernel representation. They also demonstrate that the center-support geometry need not coincide with the physical domain: the nonconvex boundary enters through the physical constraint functionals, while the population kernel is defined on the surrounding embedding region.
Singularly perturbed problem
Figure 3 illustrates the predictor–residual–corrector behavior for the singularly perturbed boundary-value problem with . The coarse predictor captures the outer solution but under-resolves the narrow boundary layer near . This is reflected by the independent residual indicator, which remains negligible over most of the interval and rises sharply only in the boundary-layer region. As a result, the adaptive strategy assigns 97 blocks to the low level, one to the mid level, and only two to the high level.
The predictor uses 200 PDE collocation points and yields an RMSE of with a maximum error of . After residual-guided enrichment, the corrector uses 250 PDE points and reduces the RMSE to and the maximum error to . Thus, the adaptive corrector improves the RMSE by a factor of 3.38 and the maximum error by a factor of 2.80 relative to a budget-matched uniform discretization using the same 250 PDE points. The uniform reference remains essentially unchanged from the coarse predictor, with an RMSE of , showing that the improvement is due to where the additional constraints are placed rather than simply to their number.
The adaptive set uses only 10.42% of the fully enriched budget obtained by assigning 24 PDE points to every block. The predictor kernel assembly and spectral solve require approximately 0.047 and 0.030 s, respectively, while residual probing and level assignment require about 0.191 s in the recorded run. The adaptive corrector requires approximately 0.090 s for kernel assembly and 0.034 s for the spectral solve, giving a total CFKS corrector solve time of 0.123 s. These results show that the residual-adaptive stage concentrates additional constraints in the boundary-layer region while leaving the underlying CFKS representation unchanged.
For context, the same singularly perturbed steady advection–diffusion problem with was considered by De Florio et al. [24]. Their experiments used interior collocation points for all methods, while the accurate ELM-based PIELM and X-TFC configurations also employed hidden neurons at this parameter value. In contrast, the present CFKS predictor uses only 200 PDE collocation points, and the residual-adaptive corrector increases this number to 250 by concentrating additional points near the localized boundary layer. Thus, rather than increasing the global representation and collocation budgets, CFKS exploits the independently evaluated residual to allocate resolution only where it is needed.

Figure 3. CFKS solution of the singularly perturbed problem. Top row: coarse predictor over the full domain, boundary-layer zoom, and independent residual indicator. Bottom row: adaptive corrector over the full domain, boundary-layer zoom, and residual-guided collocation enrichment. The adaptive refinement is concentrated near the boundary layer at .
Advection and advection–diffusion
We first consider two time-dependent transport problems to examine the performance of sequential space–time-slab marching and residual-adaptive collocation. The governing equations, computational domains, and reference solutions are summarized in Table 1. Both experiments employ a deterministic coarse predictor followed by a residual-guided corrector, with the continuum-feature mean kernel retained between the two stages. Independent residual evaluations are used to identify regions of under-resolution and redistribute the PDE collocation points accordingly. Figure 4 and Figure 5 illustrate the resulting predictor and corrector solutions, residual distributions, and space–time collocation patterns.

Figure 4. CFKS results for the linear advection problem. Top row: predictor solutions at representative times and the independently evaluated PDE residual. Bottom row: adaptive corrector solutions and the resulting space–time collocation enrichment. The enriched collocation region follows the trajectory of the advecting Gaussian pulse.

Figure 5. CFKS results for the advection–diffusion problem. Top row: predictor solutions at representative times and the independently evaluated PDE residual. Bottom row: adaptive corrector solutions and the corresponding space–time collocation enrichment. The enriched region follows the translating and diffusing Gaussian pulse.
For the linear advection problem, a narrow Gaussian pulse propagates periodically across the spatial domain with unit advection speed. The coarse predictor captures the overall motion of the pulse but accumulates visible phase and amplitude errors during propagation. At the final time , the predictor has a root-mean-square error (RMSE) of and a maximum absolute error of . The independently evaluated residual identifies a narrow moving region of under-resolution, producing a high-attention space–time band aligned with the pulse trajectory, as shown in Figure 4. Residual-guided collocation enrichment follows this band and reduces the final RMSE to and the maximum absolute error to . The corrector uses an average of 307.0 PDE collocation points per slab, with a minimum of 288 and a maximum of 320 points. The improved agreement with the exact solution demonstrates that localized enrichment can resolve the transported feature without changing the prescribed continuum-feature kernel.
The computational timings distinguish the cost of the kernel solution from the additional operations required for residual diagnostics and sequential transfer. For the advection predictor, kernel assembly, spectral factorization, and coefficient solution over all slabs require approximately 0.085, 0.119, and 0.054 s, respectively, giving a core CFKS time of 0.259 s. Residual probing and slab prediction/transfer require approximately 19.48 and 12.70 s, respectively, and therefore dominate the predictor-stage computational cost. For the adaptive corrector, the core CFKS time increases to 4.64 s, including 0.103 s for full operator-kernel assembly and 4.30 s for cached spectral factorizations associated with 52 distinct refinement patterns. The total Stage-2 elapsed time, including prediction and transfer, is 20.16 s. These measurements indicate that the cost of independent residual evaluation and dense solution transfer can substantially exceed that of the underlying population-kernel coefficient solve.
The advection–diffusion problem considers a translating Gaussian pulse on , with advection coefficient and diffusion coefficient . In contrast to pure advection, the exact solution broadens and decreases in amplitude as diffusion acts on the moving pulse. The coarse predictor reproduces the overall translation and spreading but increasingly underestimates the pulse amplitude, yielding a final RMSE of and a maximum absolute error of . As shown in Figure 5, the independent residual again identifies the moving region of strongest local variation, while the enriched collocation band broadens consistently with the diffusive spreading of the solution. The adaptive corrector reduces the final RMSE to and the maximum absolute error to . The corrector employs an average of 326.2 PDE points per slab, with the number varying between 304 and 344 according to the residual-guided refinement pattern.
For the advection–diffusion predictor, kernel assembly, spectral factorization, and coefficient solution require approximately 0.061, 0.047, and 0.048 s, respectively, resulting in a core CFKS time of 0.156 s. Residual probing requires 31.98 s, while slab prediction and transfer require 20.13 s. The adaptive corrector has a core CFKS time of 3.10 s, of which 2.69 s is spent on cached spectral factorizations for 52 distinct refinement patterns. The total Stage-2 elapsed time is 30.80 s. Together with the pure advection results, these measurements show that the numerical cost of the slab-based computations is influenced strongly by residual diagnostics, prediction-grid evaluation, and solution transfer, rather than solely by assembly and solution of the operator-kernel systems.
Taken together, the two transport experiments demonstrate that residual-adaptive CFKS can concentrate collocation resolution along moving localized structures while preserving the underlying continuum-feature representation. In the pure advection case, the enrichment follows the narrow trajectory of the periodically transported pulse, whereas in the advection–diffusion case it adapts to a translating and broadening profile. The accuracy improvements are obtained by changing the locations and number of discrete PDE constraints rather than introducing additional finite Gaussian-RBF features or sampling new feature centers. This distinction is important because the adaptive procedure operates on the discretization of the governing equations while the analytically prescribed mean kernel remains unchanged. The results also indicate that improving the efficiency of residual probing and slab-to-slab transfer is a relevant direction for reducing the overall computational cost of time-dependent CFKS implementations.
Closely related sharp-gradient transport benchmarks were previously used to motivate the distributed physics-informed extreme learning machine (PIELM) formulation [21], in which spatial domain decomposition was introduced to improve the resolution of localized solution structures. In contrast, the present CFKS computations retain a spatially global continuum-feature kernel representation within each time slab, without introducing artificial spatial interface conditions. Localized structures are instead resolved by redistributing PDE collocation points according to independently evaluated residual information. The resulting calculations illustrate how deterministic population-kernel representations can be combined with adaptive collocation without requiring spatial domain decomposition.
Viscous Burgers equation
We next consider the nonlinear viscous Burgers equation to examine the applicability of CFKS beyond linear differential operators. The problem is solved directly at viscosity over , with initial condition and homogeneous Dirichlet conditions at both spatial boundaries. The numerical solution is compared with a Cole–Hopf reference solution. At this viscosity, nonlinear steepening produces a sharp internal transition near , providing a demanding test of the continuum-feature representation, analytical kernel derivatives, and nonlinear residual minimization. The calculation employs a two-stage predictor–corrector strategy combining prescribed multiscale kernel refinement with residual-guided collocation enrichment. Unlike the linear transport experiments, both the collocation distribution and the Gaussian kernel widths are modified between the two stages.
The temporal interval is divided into 40 uniform space–time slabs, each of duration , while the spatial interval is partitioned into 20 equal blocks for residual assessment and collocation refinement. On each slab, the analytical trial-function embedding in (32) preserves the incoming solution profile and homogeneous boundary conditions. After completing a slab, the terminal solution is evaluated on a uniform spatial transfer grid containing 1201 points, with the prescribed boundary values imposed at the endpoints. A cubic-spline interpolant of this terminal profile supplies the incoming solution and its spatial derivatives for the following slab. The predictor and corrector are marched separately through the complete temporal interval, each starting from the prescribed initial condition. The predictor residual information determines the corrector collocation distribution, but the predictor solution is not imposed as the incoming profile of the corrector.
The nonlinear approximation uses the normalized global–local continuum-feature mean kernel defined in (29), with prescribed mixing weight . The global Gaussian center distribution covers the spatial interval , whereas the local component uses the narrower interval to provide additional resolution around the developing internal transition. The corresponding local-time center intervals are for the global component and for the local component. Each kernel component is normalized by its diagonal value at the spatial origin and the midpoint of the local time slab. For the Stage-1 predictor, the prescribed spatial and temporal Gaussian widths are for the global kernel and for the local kernel. In Stage 2, these widths are reduced to and , respectively. The prescribed kernel parameters remain fixed across the slabs within each stage but differ between the predictor and corrector computations.
The nonlinear coefficient solution on each slab begins with a frozen-advection linearization based on the incoming solution profile, followed by a Picard update. The linearized coefficient systems are solved using singular-value-decomposition (SVD) regularization, with relative ridge factors of and for the frozen-advection and Picard steps, respectively. The resulting coefficients initialize nonlinear least-squares minimization of the full Burgers residual, using the analytical Jacobian derived in (36). The nonlinear solver employs a trust-region reflective algorithm with an iterative least-squares linear solver, a maximum of 20 function evaluations, and convergence tolerances of for the objective function, coefficient changes, and gradient norm. This initialization and optimization procedure is used in both stages, so the distinction between the predictor and corrector lies principally in their prescribed kernel scales and collocation distributions.
For the Stage-1 predictor, each of the 20 spatial blocks contains 10 PDE collocation points, giving 200 points per time slab. After completing the predictor march, the nonlinear PDE residual is evaluated on independent probe grids within each block. Each probe grid contains 11 spatial locations and 7 local-time locations, giving 77 residual evaluation points per block. The spatial probe locations extend from 5% to 95% of the block width, while the temporal locations range from to . Residual root-mean-square and maximum magnitudes are evaluated separately for each block, and the normalized indicator in (40) is computed using and . The normalization is performed independently within each time slab, allowing the refinement decisions to respond to the spatial distribution of residuals at different stages of the evolution.
The refinement thresholds are prescribed as and , defining low-, medium-, and high-attention regions. To avoid abrupt changes in collocation density, medium- and high-attention assignments are extended to their immediate neighboring spatial blocks without wrapping around the physical boundaries. High-attention assignments take precedence wherever the expanded regions overlap. The corresponding collocation budgets are , , and PDE points per block, with deterministic collocation locations generated using a two-dimensional Halton sequence with bases 2 and 3. Within each block, the selected number of Halton points is mapped to the corresponding spatial and local-time interval. Since the kernel representer sites coincide with the PDE collocation locations, residual-guided enrichment changes both the number of PDE constraints and the dimension of the nonlinear coefficient system. The resulting Stage-2 corrector uses between 350 and 600 PDE points per slab, with an average of 402.75 points.
Figure 6 illustrates the two-stage solution process and its associated collocation refinement. The coarse predictor initially follows the Cole–Hopf reference solution but develops oscillatory errors as the nonlinear internal transition becomes sharper. These discrepancies are especially pronounced near the central gradient at later times, where the independently evaluated PDE residual also identifies increased under-resolution. The Stage-2 corrector allocates additional collocation points around the developing transition while using the narrower prescribed multiscale kernel widths. The corrected solution exhibits substantially improved agreement with the reference, indicating that the combined refinement strategy can resolve localized nonlinear structures more accurately than the coarse predictor.

Figure 6. Two-stage multiscale CFKS solution of viscous Burgers’ equation at . Top row: coarse predictor solutions at , , and , together with the independently evaluated nonlinear PDE residual. Bottom row: adaptive corrector solutions at the corresponding times and the residual-guided collocation distribution. Dashed curves denote the Cole–Hopf reference solution. The corrector combines increased collocation resolution near the developing internal transition with narrower prescribed multiscale kernel widths.
The quantitative errors at four representative times are summarized in Table 4. At , the predictor RMSE is , which decreases to after correction. As the internal transition develops, the difference between the two stages becomes considerably more pronounced. At , the predictor RMSE of is reduced to , while at the corresponding values are and . At the final time , the predictor has an RMSE of and a maximum absolute error of , whereas the corrector reduces these values to and , respectively. The final-time RMSE is therefore reduced by approximately a factor of 106.
| Predictor | Corrector | |||
| RMSE | RMSE | |||
| 0.25 | ||||
| 0.50 | ||||
| 0.75 | ||||
| 1.00 |
Table 4. Predictor and corrector errors for viscous Burgers’ equation at , evaluated against the Cole–Hopf reference solution.
Figure 7 provides a further comparison between the corrected CFKS solution and the Cole–Hopf reference at , 0.5, 0.75, and 1. The corrected approximation remains in close agreement with the reference throughout the simulation, including the increasingly steep internal transition near . Both the amplitude and position of this localized structure are reproduced without spatial domain decomposition or viscosity continuation. The improvement should, however, be interpreted as the combined effect of residual-guided collocation enrichment and refinement of the prescribed Gaussian kernel widths, since these two changes are introduced simultaneously in the corrector.

Figure 7. Corrected multiscale CFKS solution of viscous Burgers’ equation at compared with the Cole–Hopf reference solution at , , , and . The corrected solution closely follows the reference while resolving the developing internal transition near .
Overall, the Burgers experiment demonstrates that the continuum-feature mean-kernel representation can be extended to nonlinear PDEs through linearized initialization and nonlinear least-squares residual minimization. Analytical kernel derivatives provide the nonlinear PDE residual and its Jacobian, while the global–local multiscale representation supplies additional flexibility for resolving localized solution structures. The two-stage procedure improves accuracy by combining residual-guided collocation enrichment with prescribed kernel-scale refinement. The computation is performed directly at the target viscosity without introducing finite Gaussian-RBF center sampling, viscosity continuation, or spatial domain decomposition.
Finite-feature to continuum convergence
To verify the finite-feature interpretation of CFKS, we use the first-order oscillatory ODE benchmark defined in § 3,
with exact solution . This case provides a convenient controlled setting in which the finite Gaussian-RBF feature representation and the corresponding CFKS population limit can be compared directly.
Figure 8 verifies numerically that the finite Gaussian-RBF feature representation approaches the CFKS population limit as the number of sampled features increases. The top panel shows the relative Frobenius errors of the empirical value Gram matrix, , and of the operator-applied Gram matrix, . Both decrease systematically with and closely follow the reference slope expected for Monte Carlo sampling of the center population. The mean value-kernel error decreases from 1.52 at to at , while the operator-Gram error decreases from 1.74 to over the same range. The shrinking realization-to-realization spread further confirms convergence toward a deterministic population kernel.

Figure 8. Finite-feature convergence toward CFKS. Left: relative errors of the empirical value and operator Gram matrices as the number of sampled Gaussian RBF features increases; shaded bands denote realization-to-realization variability and the dashed line indicates the reference rate. Right: RMS difference between the finite-feature and CFKS solutions, together with the finite-feature error relative to the exact solution. The dashed horizontal line denotes the CFKS error relative to the exact solution.
The bottom panel shows the corresponding effect at the solution level. For small feature counts, the finite-feature approximation remains far from the CFKS solution, with RMS differences of approximately for . A marked transition occurs once the feature space becomes sufficiently rich: at the mean finite-feature-to-CFKS RMSE falls to , and subsequently decreases to at . The CFKS solution itself has an RMSE of relative to the exact solution. At large , the finite-feature solutions therefore converge toward the deterministic CFKS reference, while their realization-dependent variability also decreases.
These results provide direct numerical evidence for the interpretation of CFKS as the population limit of the underlying finite Gaussian-RBF feature model. The continuum formulation removes the finite-feature sampling error and the associated realization variability, while retaining the operator Gram structure obtained from the original RBF representation.
Conditioning and regularization
Figure 9 examines the sensitivity of CFKS to the two principal numerical parameters using the first-order oscillatory ODE benchmark. The regularization sweep shows the expected trade-off between stabilization and approximation accuracy: insufficient regularization leaves the solve sensitive to small singular values, whereas excessive regularization increases both the solution and residual errors. The baseline value lies within the stable low-error region.

Figure 9. Conditioning and regularization study for the first-order oscillatory ODE benchmark. Left: dependence of the CFKS solution RMSE and PDE-residual RMS on the relative Tikhonov parameter at fixed . Right: raw reduced-system condition number and regularized effective condition number as the Gaussian width is varied. Dashed vertical lines indicate the baseline parameters used in the numerical experiments.
The lower panel shows that the Gaussian width has a strong effect on the conditioning of the reduced operator system. Increasing changes the overlap among neighboring population-kernel basis functions and can substantially increase the raw condition number. Tikhonov regularization reduces this effective conditioning burden while preserving the solution accuracy over a useful range of kernel widths. These results motivate the moderate width and weak regularization used throughout the remaining experiments.
Advection with two separated Gaussian pulses
To further test the spatial selectivity of the residual-adaptive procedure, we consider the same periodic advection equation as in § 3, but replace the single Gaussian initial condition with two separated pulses,
The two pulses therefore have identical width and are initially centered at and . The exact periodic solution is obtained by translating the complete initial profile with unit advection speed.
Figure 10 shows that the predictor captures the transport of both pulses but develops localized oscillatory errors near the moving sharp features. The independent residual identifies two distinct space–time trajectories associated with the advecting pulses, and the corrector consequently concentrates additional collocation points along both paths. The final-time predictor error is RMSE with a maximum error of . After residual-guided enrichment, these errors decrease to RMSE and a maximum error of , respectively.

Figure 10. Periodic advection of two separated Gaussian pulses. Top row: predictor solution at , , and , followed by the independently evaluated residual over the space–time domain. Bottom row: adaptive-corrector solution at the same times and the resulting collocation-enrichment pattern. The residual traces both propagating pulses and directs increased collocation density along their corresponding space–time trajectories.
The adaptive corrector uses an average of 413.0 PDE collocation points per slab, with a range of 384–432. Thus, the residual-based enrichment is not tied to a single localized feature; it can identify and track multiple simultaneously propagating structures while leaving the underlying CFKS kernel and physical constraints unchanged.
Overall observations
The numerical experiments demonstrate the applicability of CFKS across differential problems involving oscillatory solutions, multiscale structures, irregular geometries, higher-order operators, localized boundary layers, time-dependent transport, and nonlinear dynamics. The finite-feature convergence study provides direct numerical evidence that empirical Gaussian-RBF value and operator Gram matrices approach their analytically defined continuum counterparts as the number of sampled features increases. This establishes the connection between finite Gaussian-RBF representations and the deterministic continuum-feature formulation used throughout the paper. The ODE, Poisson, and biharmonic benchmarks further demonstrate that the resulting mean kernel and its analytical derivatives can be employed for differential operators of different orders without introducing separate finite feature realizations. The successful treatment of the irregular five-petal domain also illustrates that the prescribed kernel center-support region need not coincide with the physical computational domain, since the domain geometry and physical conditions enter through the corresponding constraint functionals.
The singularly perturbed and time-dependent transport experiments demonstrate the role of residual-adaptive collocation as an additional resolution strategy within the continuum-feature framework. Independent PDE residual evaluations identify localized regions where the predictor approximation is insufficiently resolved, and the corrector redistributes collocation points accordingly. The singularly perturbed problem shows that concentrating additional constraints near a narrow boundary layer can be more effective than uniform refinement with the same point budget. Similarly, the advection and advection–diffusion experiments demonstrate that residual-guided enrichment can follow moving localized structures, while the two-pulse advection benchmark confirms that multiple separated regions of under-resolution can be identified simultaneously. In these linear examples, the prescribed continuum-feature mean kernel remains unchanged between the predictor and corrector stages; refinement modifies the discrete collocation representation rather than the underlying Gaussian feature distribution.
The viscous Burgers experiment extends the numerical assessment to a nonlinear differential operator. In this setting, the continuum-feature mean kernel defines a deterministic trial representation, while the unknown coefficients are determined through nonlinear least-squares residual minimization. Linearized initialization and analytical kernel derivatives provide the starting approximation, nonlinear residual, and corresponding Jacobian. At the investigated viscosity , the two-stage procedure substantially improves the resolution of the developing internal transition.
Unlike the linear transport experiments, however, the Burgers corrector combines residual-guided collocation enrichment with narrower prescribed Gaussian widths in a global–local multiscale kernel. The observed accuracy improvement therefore reflects their combined effect, and the present experiment does not separately quantify the contribution of either refinement mechanism.
The conditioning and timing results also clarify the computational characteristics of CFKS. The regularization and Gaussian-width study for the first-order ODE confirms that the accuracy and conditioning of the discrete operator-kernel system depend on these prescribed numerical parameters. In the steady linear benchmarks, the main computational costs are associated with analytical kernel assembly and dense linear algebra, while null-space construction contributes comparatively little to the measured solve times. For the slab-based transport problems, residual evaluation, dense solution prediction, and temporal transfer can require more computational time than the core kernel coefficient solves. Nonlinear problems introduce additional costs through repeated residual and Jacobian evaluations during iterative coefficient optimization. Consequently, the overall computational efficiency depends on both the kernel discretization and the numerical procedures used for refinement, diagnostics, and time advancement.
Taken together, these findings support CFKS as a deterministic computational framework connecting finite Gaussian-RBF feature representations with analytically constructed continuum-feature kernel PDE solvers. Its principal representational advantage is the elimination of the finite feature count and individual center-sampling variability, rather than the elimination of all numerical parameter choices. Gaussian widths, bounded center distributions, collocation configurations, regularization, and, where applicable, nonlinear initialization remain important aspects of the numerical formulation. The experiments demonstrate the flexibility of this approach across the considered benchmarks, while also identifying parameter selection, conditioning, and computational scalability as important considerations for broader applications.
Remarks
Scope of nonlinear validation. The nonlinear CFKS formulation is demonstrated numerically using viscous Burgers’ equation at . Although the analytical kernel derivatives and nonlinear least-squares formulation are applicable to other sufficiently smooth nonlinear differential operators, their numerical accuracy, robustness, and convergence behavior have not been systematically investigated across broader classes of nonlinear PDEs. The reported Burgers corrector modifies both the collocation distribution and the prescribed multiscale kernel widths, so their individual effects cannot be distinguished from the present results. Further numerical studies involving different nonlinear operators, parameter regimes, and separate refinement strategies are needed to assess the generality and limitations of the nonlinear extension.
Kernel parameters and numerical stability. Although analytical marginalization removes finite-feature sampling variability, CFKS remains dependent on prescribed Gaussian widths, center distributions, and regularization parameters. The conditioning study illustrates the influence of Gaussian width and regularization on a representative linear problem, but does not establish a universally optimal parameter-selection rule. For problems containing localized structures, the appropriate kernel scales must also be considered in relation to the spatial and temporal resolution. In sequential space–time-slab computations, the accuracy of transferring the terminal solution between slabs provides an additional numerical consideration. Systematic parameter-selection procedures and conditioning-aware kernel refinement therefore remain worthwhile directions for further investigation.
Computational scalability. CFKS avoids explicit finite Gaussian-RBF feature realizations but retains the computational demands of dense kernel discretizations. For a linear system involving constraints, storing the full operator Gram matrix requires memory, while direct dense factorization may require operations. Nonlinear formulations additionally require repeated evaluation of the PDE residual and its Jacobian. Sequential space–time-slab marching limits the size of individual temporal problems, and residual-adaptive collocation can concentrate resolution in selected regions, but neither mechanism removes the general scalability limitations of dense kernel systems. Low-rank kernel approximations, matrix-free iterative methods, efficient residual evaluation, and localized or domain-decomposed formulations are potential approaches for extending CFKS to larger-scale applications.
Conclusions
This work introduced Continuum-Feature Kernel Solver (CFKS), a deterministic framework for solving differential equations through analytical marginalization of a bounded continuum of Gaussian radial basis function (RBF) features. By replacing a finite feature realization with its exact continuum-feature mean kernel, CFKS eliminates the need to prescribe a finite feature count or select individual feature centers. The resulting kernel and its differential-operator counterparts are evaluated analytically and used directly in the numerical discretization. For linear problems, this construction leads to symmetric operator-kernel collocation systems, with boundary and initial conditions enforced through constraint-compatible null-space parameterization. For nonlinear problems, the continuum-feature representation is retained, while the unknown coefficients are determined through linearized initialization and nonlinear least-squares residual minimization using analytical kernel derivatives and Jacobians.
The numerical investigations support the proposed formulation from both the feature-representation and differential-equation perspectives. Finite-feature convergence experiments demonstrate that empirical Gaussian-RBF value and operator Gram matrices approach their analytically derived continuum counterparts as the number of sampled features increases. The corresponding finite-feature solutions also approach the deterministic CFKS solution, confirming the computational connection between finite Gaussian-RBF models and the continuum-feature formulation. The linear benchmarks demonstrate the applicability of the analytical operator kernels to oscillatory and multiscale ODEs, second- and fourth-order elliptic problems on irregular domains, singularly perturbed boundary layers, and time-dependent transport equations. Residual-guided collocation enrichment further improves the resolution of localized solution structures without changing the underlying mean kernel in the linear adaptive examples. The transport experiments additionally demonstrate that this strategy can track multiple moving localized features within a spatially global kernel formulation.
The viscous Burgers experiment extends the numerical assessment to a nonlinear differential operator. At the investigated viscosity of , the continuum-feature mean kernel, analytical constraint embedding, and nonlinear residual minimization provide an accurate approximation of the developing internal transition. The two-stage predictor–corrector procedure combines residual-guided collocation enrichment with narrower prescribed global and local Gaussian kernel widths, producing substantial improvements over the coarse predictor. These results demonstrate the feasibility of extending CFKS to nonlinear PDEs without introducing finite Gaussian-RBF feature realizations. However, because both the collocation distribution and kernel scales are modified between stages, the individual contributions of these refinement mechanisms cannot be isolated from the present experiment. Further assessment across different nonlinear operators and parameter regimes is therefore needed.
The principal significance of CFKS lies in using the continuum-feature limit as a direct computational representation rather than solely as an asymptotic interpretation of a finite feature model. Analytical marginalization removes finite-feature sampling variability while preserving the connection to the underlying Gaussian-RBF representation. Nevertheless, the method retains prescribed Gaussian widths, bounded center distributions, collocation configurations, and regularization parameters, whose selection influences accuracy and conditioning. Moreover, CFKS inherits the dense matrix assembly and linear-algebra costs of global kernel methods, while nonlinear problems require additional iterative computations. The time-dependent experiments also highlight the computational expense associated with residual evaluation and slab-to-slab solution transfer. Future research will address scalable low-rank and iterative kernel formulations, efficient residual evaluation, conditioning-aware parameter selection, and localized or domain-decomposed extensions. Broader nonlinear benchmarks and systematic investigations of collocation and kernel-scale refinement will further clarify the accuracy, robustness, and computational limits of CFKS.
From finite features to the CFKS representer form
We briefly show how the CFKS representer form follows from the finite RBF feature model. Consider
and let denote the selected PDE, boundary, and initial functionals. Define the finite feature matrix
and the feature vector
Then
For a minimum-norm feature-space solution satisfying the discrete constraints, the coefficient vector lies in the range of and may therefore be written as
for some coefficient vector . Substituting into the finite feature representation gives
Because acts on the second argument, the term in brackets can be written as
where the empirical feature kernel is
Hence
If the feature parameters are sampled from , then as ,
Under the regularity conditions used in this work, the linear functionals may be interchanged with the expectation, so that
The continuum-limit representation is therefore
which is the CFKS representer form used in the main text.
This derivation also shows that the finite feature coefficients disappear from the continuum formulation: their effect is transferred to the constraint-space coefficients through the empirical Gram structure, while the sampled kernel is replaced by the analytically evaluated continuum-feature mean kernel .
Analytical Derivation of the Bounded Continuum-Feature Gaussian Kernel
This appendix derives the closed-form mean kernel associated with a continuum of Gaussian radial-basis-function features whose centers are uniformly distributed over a bounded interval. The full-line version of this construction is classical and yields a squared-exponential covariance; see, for example, Rasmussen and Williams [8]. Here we retain the finite center-support interval explicitly, which produces additional error-function terms.
One-dimensional Gaussian feature model
Consider the Gaussian feature
where the center is distributed uniformly over
The corresponding mean feature kernel is
Substituting eq. 64 gives
The integral can be evaluated analytically by completing the square with respect to the center variable .
Completion of the square
Define the midpoint and separation coordinates
We first rewrite the quadratic form in the exponent:
Using
and
we obtain
Substituting eq. (72) into eq. (67) gives
Thus, the kernel separates into a Gaussian factor depending on the relative distance and a finite-support correction depending on the midpoint .
Evaluation using the error function
Introduce the change of variables
The limits become
Therefore,
Using the error function
we have
Hence,
Substitution into eq. (73) yields the closed form
Equation eq. (80) is the bounded continuum-feature Gaussian kernel used in this work.
Interpretation of the bounded-support correction
The expression in eq. (80) can be written schematically as
where
depends only on the point separation, while
depends on the midpoint .
This decomposition shows explicitly why the bounded-center kernel is generally nonstationary. If the center support is finite, two pairs of points with the same separation can have different covariance if their midpoints lie at different distances from the center-support boundaries.
Away from the edges and , and provided the support is wide relative to , the two error functions approach and , respectively, so that the support correction becomes approximately constant. Near the edges, however, the finite-support correction modifies the kernel.
Full-line limit
The relationship with the standard Gaussian or squared-exponential kernel is obtained by taking
Since
the bracketed factor in eq. (80) tends to 2.
If one considers the unnormalized center measure
rather than the uniform probability density , the corresponding full-line feature integral is
After normalization of the amplitude, this is a squared-exponential kernel with effective length scale
under the conventional form
Thus, the usual stationary Gaussian kernel is recovered as the infinite-support limit of the Gaussian center-feature construction, while the bounded-center model retains additional support information.
Positive semidefiniteness
The positive semidefiniteness of eq. (66) follows immediately from its feature representation. For arbitrary points and coefficients ,
Hence, the continuum-center construction defines a valid positive semidefinite kernel independently of the closed-form calculation.
Differentiation of the mean kernel
For PDE applications, the essential quantity is not only itself but its derivatives with respect to both arguments. Since the Gaussian feature and all of its derivatives are smooth, and the center support is bounded, differentiation may be interchanged with integration:
Equation (91) gives a direct mean interpretation of every differentiated kernel block.
For example,
and
Therefore,
and similarly,
For a differential operator ,
Thus, the operator kernel used in the continuum-feature collocation method is precisely the mean Gram kernel of the operator-applied Gaussian features.
Gaussian moment representation
Higher-order derivatives can be evaluated efficiently without repeatedly differentiating the error-function expression eq. (80). After completing the square, every mixed derivative reduces to finite combinations of Gaussian moments of the form
The first moments are
and
For , integration by parts gives the recursion
This recursion provides a convenient analytic route to the high-order kernel derivatives required by elliptic, advection–diffusion, and biharmonic operators.
Extension to multiple dimensions
Suppose the -dimensional Gaussian feature is separable,
and the center coordinates are independently distributed as
The mean density factorizes,
and therefore the multidimensional kernel also factorizes:
Consequently, mixed derivatives also factorize:
This property is particularly useful for Cartesian embedding domains and space–time formulations. For example, a two-dimensional Laplacian kernel can be assembled from one-dimensional derivative factors, while space–time advection–diffusion kernels can be constructed from separate spatial and temporal mean kernels.
Summary
The continuum-center Gaussian feature model therefore admits the exact bounded-support kernel
which reduces to a Gaussian/squared-exponential kernel in the full-support limit after normalization.
More importantly for PDE collocation, the differentiated mean kernel satisfies
Hence, the differential-operator kernel matrices used by the proposed method are exactly the continuum limits of the corresponding finite Gaussian-RBF operator Gram matrices.
CRediT authorship contribution statement
Vikas Dwivedi: Conceptualization, Methodology, Software, Writing - Original Draft, Bruno Sixou and Monica Sigovan: Supervision and Writing - Review & Editing.
Acknowledgements
This work was supported by the ANR (Agence Nationale de la Recherche), France, through the RAPIDFLOW project (Grant no. ANR-24-CE19-1349-01).
References
- [1]M. Raissi, P. Perdikaris, and G.E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019.DOI
- [2]E.J. Kansa. Multiquadrics—a scattered data approximation scheme with applications to computational fluid-dynamics—i surface approximations and partial derivative estimates. Computers & Mathematics with Applications, 19(8):127–145, 1990.DOI
- [3]E.J. Kansa. Multiquadrics—a scattered data approximation scheme with applications to computational fluid-dynamics—ii solutions to parabolic, hyperbolic and elliptic partial differential equations. Computers & Mathematics with Applications, 19(8):147–161, 1990.DOI
- [4]C. Franke and R. Schaback. Solving partial differential equations by collocation using radial basis functions. Applied Mathematics and Computation, 93(1):73–82, 1998.DOI
- [5]Y.C. Hon and R. Schaback. On unsymmetric collocation by radial basis functions. Applied Mathematics and Computation, 119(2):177–186, 2001.DOI
- [6]B. Jumarhon, S. Amini, and K. Chen. The hermite collocation method using radial basis functions. Engineering Analysis with Boundary Elements, 24(7):607–611, 2000.DOI
- [7]Christopher Williams. Computing with infinite networks. In M.C. Mozer, M. Jordan, and T. Petsche, editors, Advances in Neural Information Processing Systems, volume 9. MIT Press, 1996.
- [8]Christopher KI Williams and Carl Edward Rasmussen. Gaussian processes for machine learning, volume 2. MIT press Cambridge, MA, 2006.DOI
- [9]Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines. In J. Platt, D. Koller, Y. Singer, and S. Roweis, editors, Advances in Neural Information Processing Systems, volume 20. Curran Associates, Inc., 2007.
- [10]Jingrun Chen, Xurong Chi, Weinan E, and Zhouwang Yang. Bridging traditional and machine learning-based algorithms for solving pdes: The random feature method, 2022.arxiv.org/abs/2207.13380
- [11]Jiale Linghu and Yangshuai Wang. Residual-christoffel sampling for random feature collocation of linear pdes, 2026.arxiv.org/abs/2607.13382
- [12]Siran Chen and Wenzhong Zhang. Solving partial differential equations with analytic random feature method. East Asian Journal on Applied Mathematics, 17(2):274–292, 2027.DOI
- [13]Thore Graepel. Solving noisy linear operator equations by gaussian processes: Application to ordinary and partial differential equations. In Proceedings of the Twentieth International Conference on Machine Learning, January 2003.
- [14]Jon Cockayne, Chris Oates, Tim Sullivan, and Mark Girolami. Probabilistic numerical methods for pde-constrained bayesian inverse problems. AIP Conference Proceedings, 1853(1):060001, 06 2017.DOI
- [15]David Dalton, Alan Lazarus, Hao Gao, and Dirk Husmeier. Boundary constrained gaussian processes for robust physics-informed machine learning of linear partial differential equations. Journal of Machine Learning Research, 25(272):1–61, 2024.
- [16]Davide Elia De Falco, Enrico Schiassi, and Francesco Calabrò. Least squares with equality constraints extreme learning machines for the resolution of pdes. Journal of Computational Physics, 547:114553, 2026.
- [17]Rishi Mishra, Anuj Said, Smriti Smriti, Ganapathy Krishnamurthi, Balaji Srinivasan, and Sundararajan Natarajan. A unified null space projected physics-informed extreme learning machine framework for static bending of kirchhoff–love and reissner–mindlin plates. Computer Methods in Applied Mechanics and Engineering, 461:119172, 2026.DOI
- [18]Rishi Mishra, Smriti, Ganapathy Krishnamurthi, Balaji Srinivasan, and Sundararajan Natarajan. Eig-pielm: an efficient mesh-free method for eigenvalue problems using physics-informed extreme learning machines. Computer Methods in Applied Mechanics and Engineering, 451:118674, 2026.
- [19]Chenxi Wu, Min Zhu, Qinyang Tan, Yadhu Kartha, and Lu Lu. A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 403:115671, 2023.DOI
- [20]Lu Lu, Xuhui Meng, Zhiping Mao, and George Em Karniadakis. Deepxde: A deep learning library for solving differential equations. SIAM Review, 63(1):208–228, 2021.DOI
- [21]Vikas Dwivedi, Nishant Parashar, and Balaji Srinivasan. Distributed learning machines for solving forward and inverse problems in partial differential equations. Neurocomputing, 420:299–316, 2021.DOI
- [22]Sifan Wang, Shyam Sankaran, and Paris Perdikaris. Respecting causality for training physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 421:116813, 2024.DOI
- [23]Ben Moseley, Andrew Markham, and Tarje Nissen-Meyer. Finite basis physics-informed neural networks (fbpinns): a scalable domain decomposition approach for solving differential equations. Advances in Computational Mathematics, 49(4):62, Jul 2023.DOI
- [24]Mario De Florio, Enrico Schiassi, Francesco Calabrò, and Roberto Furfaro. Physics-informed neural networks for 2nd order odes with sharp gradients. Journal of Computational and Applied Mathematics, 436:115396, 2024.DOI