Learning-guided Kansa Collocation for Forward and Inverse PDE Problems¶

Conference: ICLR 2026 arXiv: 2602.07970 Code: None Area: Scientific Computing / PDE Solving Keywords: Kansa method, radial basis functions, nonlinear PDEs, inverse problems, neural PDE solvers

TL;DR¶

This work extends the meshfree radial basis function (RBF)-based Kansa collocation method from single-variable linear PDEs to coupled multi-variable and nonlinear PDE settings. It incorporates automatic shape-parameter tuning and multiple time-stepping schemes, and provides a systematic comparison against neural PDE solvers such as PINNs and FNO on both forward and inverse problems.

Background & Motivation¶

Challenges in PDE solving: Partial differential equations are widely used in physics, graphics, and biology, yet classical numerical methods (FDM/FEM) suffer from the curse of dimensionality, high computational cost, and domain-specific discretization requirements.

Rise of neural PDE solvers: PINNs (Raissi et al. 2019) and FNO (Li et al. 2020) demonstrate strong generalization and high-dimensional handling capabilities, but are limited by high training costs and large data requirements.

Advantages of the Kansa method: The Kansa method is a meshfree RBF-based solver that requires no grid discretization and is naturally suited to complex geometries. Zhong et al. (2023) introduced the Constrained Neural Fields (CNF) framework with automatic shape-parameter optimization.

Limitations of existing Kansa methods: Zhong et al. (2023) address only single-variable linear PDEs and cannot handle the coupled systems and nonlinear operators common in practice.

Lack of systematic comparison: It remains unclear how an extended Kansa method compares to classical and neural PDE solvers across multiple quality metrics (L1/L2 error, efficiency, convergence rate, etc.).

Importance of inverse problems: Inferring unknown PDE parameters (e.g., diffusion coefficients, flow velocities) from observational data is critical for scientific simulation, yet existing Kansa frameworks have not addressed the inverse setting.

Method¶

Overall Architecture¶

The framework is built on the Kansa collocation method: the field is approximated as a linear combination of RBFs, \(\hat{u}(\mathbf{x}) = \sum_k \alpha_k \psi_k(\|\mathbf{x} - \mathbf{x}_k\|)\), and the coefficients \(\alpha_k\) are determined by enforcing PDE constraints and boundary conditions. The core extensions are: (1) coupled multi-variable extension, (2) nonlinear operator handling, (3) automatic hyperparameter tuning, and (4) inverse problem solving.

Key Designs¶

Coupled Multi-variable PDE Extension (Extension 1)
Function: Extends a single unknown field \(u\) to a multi-dimensional field \(\mathbf{u} = [u_1, u_2, \ldots, u_{N_D}]\).
Mechanism: Each dimension \(u_d\) has an independent RBF expansion and coefficients \(\alpha_k^{(d)}\); a linear coupling operator \(\mathcal{G}\) links the dimensions, yielding a horizontally stacked block-matrix system.
Design Motivation: Many physical equations (e.g., Navier–Stokes, Maxwell) are inherently coupled PDE systems requiring the simultaneous solution of multiple physical quantities.
Nonlinear Operator Handling (Extension 2)
Function: Handles nonlinear differential operators such as the \(u \frac{\partial u}{\partial x}\) term in Burgers' equation.
Mechanism: Introduces the differentiation matrix \(\mathbf{D}_x = \mathbf{K}_x \cdot \mathbf{K}^{-1}\) to decompose nonlinear operators into combinations of differential operations on known fields. Five solution strategies are provided: forward Euler (explicit), IMEX (semi-implicit), backward Euler (implicit Newton–Raphson), Crank–Nicolson (second-order), and full nonlinear direct optimization.
Design Motivation: Nonlinearity prevents direct separation into a linear system; temporal discretization or iterative optimization is required to circumvent this.
Automatic Shape-Parameter Tuning
Function: Automatically optimizes the RBF shape parameter \(\epsilon\).
Mechanism: For linear PDEs, jointly minimizes the condition number of the operator matrix and the variation of the solution field. For nonlinear PDEs, directly minimizes a composite objective of PDE residual, solution variation, and training L2 loss.
Design Motivation: The choice of \(\epsilon\) critically affects accuracy and stability (accuracy–conditioning trade-off), making manual tuning impractical.
Inverse Problem Solving
Function: Recovers unknown PDE parameters \(\boldsymbol{\pi}\) from solution observations \(u^{\text{obs}}\).
Mechanism: \(\boldsymbol{\pi}^* = \arg\min_{\boldsymbol{\pi}} \mathcal{L}(u^{\text{obs}}, u^{\text{pred}}(\boldsymbol{\pi}))\), solved using SciPy's least-squares and root-finding algorithms.
Design Motivation: Inverse problems are critical in scientific computing (e.g., parameter estimation, material property inference); this extension broadens the Kansa framework to the inverse setting.
Systematic Comparison with Other Solvers
Function: Benchmarks Kansa, PINN, and FNO on standard PDE problems.
Mechanism: Unified evaluation metrics (L2 error, efficiency, memory, convergence rate) with fairly matched training data volumes.
Design Motivation: Provides a solver-selection guide for different PDE types.

Loss & Training¶

Linear PDEs: Least-squares closed-form solution \(\mathbf{a}^{\text{opt}} = (\mathbf{F}^T\mathbf{F})^{-1}\mathbf{F}^T\mathbf{h}\).
Nonlinear PDEs: Residual minimization \(\min_\alpha \sum_i (\mathcal{F}[\hat{u}](\mathbf{x}_i) - h(\mathbf{x}_i))^2\).
Shape-parameter tuning: Grid search over \(\epsilon\).
PINN baseline: Adam optimizer, learning rate \(10^{-3}\), 3000 epochs.
FNO baseline: Requires 100 PDE instances for training, 100 epochs.

Key Experimental Results¶

Main Results¶

Relative L2 error comparison on the 1D advection equation (forward problem):

Method	Type	Relative L2 Error	Notes
Kansa (linear solve)	Meshfree RBF	Low	No training; direct solve
Kansa (IMEX)	Meshfree RBF	Low, stable	Semi-implicit; suited for stiff problems
Kansa (Crank–Nicolson)	Meshfree RBF	Lowest	Second-order accuracy
PINN	Neural network	Moderate	Requires many training iterations
FNO	Operator learning	Moderate	Requires multi-instance training data

Comparison on Burgers' equation (nonlinear):

Method	Scheme	Accuracy	Stability
Forward Euler Kansa	Explicit	\(O(\Delta t)\)	Unstable
IMEX Kansa	Semi-implicit	\(O(\Delta t)\)	Stable
Backward Euler Kansa	Implicit + Newton	\(O(\Delta t)\)	Stable
Crank–Nicolson Kansa	Implicit	\(O(\Delta t^2)\)	Stable
Full nonlinear Kansa	Global optimization	\(O(1)\)	N/A

Ablation Study¶

Dimension	Variable	Observation
Number of collocation points \(N\)	50 → 500	Accuracy improves but condition number deteriorates
Shape parameter \(\epsilon\)	Manual vs. auto-tuned	Auto-tuning significantly reduces error
Time-stepping scheme	5 schemes	Crank–Nicolson achieves highest accuracy; IMEX offers best overall balance
Coupled dimensions \(N_D\)	1 → multi	Computational cost grows linearly; accuracy is largely preserved

Key Findings¶

The Kansa method substantially outperforms PINNs in accuracy at low collocation-point counts, with a clear advantage in settings that do not require large training datasets.
Automatic shape-parameter tuning effectively resolves the accuracy–conditioning trade-off inherent to RBF methods.
Among the nonlinear extensions, the IMEX scheme provides the best balance of accuracy, stability, and efficiency.
The differentiability of the Kansa method makes parameter inference in inverse problems natural and efficient.
FNO, while offering strong generalization, requires approximately 100 times more training data than Kansa or PINNs.

Highlights & Insights¶

Systematic extension: The progression from single-variable linear to coupled nonlinear PDEs is logically structured, and the five nonlinear solution strategies cover a broad range of practical requirements.
Effective use of the differentiation matrix: \(\mathbf{D}_x = \mathbf{K}_x \cdot \mathbf{K}^{-1}\) combines the flexibility of RBFs with the precision of differential operators, serving as the key enabler for extending Kansa to the nonlinear regime.
Practical orientation: The systematic solver comparison provides actionable guidance for real-world scientific computing applications.
Meshfree advantages: The Kansa method requires no mesh generation and is inherently well-suited to complex geometries and high-dimensional problems.
Natural integration of inverse problems: The RBF representation reduces parametric inversion to a standard optimization problem.

Limitations & Future Work¶

Ill-conditioning: The RBF matrix condition number deteriorates rapidly at high collocation-point densities, limiting scalability.
High-dimensional extension: Experiments are restricted to 1D and simple 2D cases; validation in 3D and higher dimensions is absent.
Convergence guarantees for nonlinear solvers: Convergence of the Newton–Raphson solver depends on initialization and lacks theoretical guarantees.
Incomplete comparison with modern methods: Comparisons with more recent approaches such as DeepONet and Operator Transformers are not conducted.
Limited inverse problem experiments: Inverse problem evaluation covers only simple parameter inference; more complex scenarios (e.g., unknown source terms, unknown boundary conditions) are not tested.
Insufficient theoretical error analysis: Error bounds and convergence-order analysis for the nonlinear extensions are not adequately developed.

Zhong et al. (2023) proposed the CNF framework that directly motivates this work; the present paper is a natural extension toward coupled and nonlinear settings.
Kansa (1990) provides the theoretical foundation for meshfree RBF methods.
Raissi et al. (2019) (PINN) and Li et al. (2020) (FNO) represent two major paradigms in neural PDE solving and serve as the primary baselines.
Insights: Integrating the Kansa method with differentiable rendering pipelines for inverse physics problems is a promising direction; hybrid approaches combining the Kansa method with neural operators (e.g., using neural networks to learn optimal collocation point placement) also merit exploration.

Rating¶

Novelty: ⭐⭐⭐ — The extension directions are natural but largely incremental; the core ideas (differentiation matrices, temporal discretization) are combinations of classical numerical techniques.
Experimental Thoroughness: ⭐⭐⭐ — Forward and inverse problems are covered across multiple PDE types, but experiments are small-scale (primarily 1D) and lack large-scale validation.
Writing Quality: ⭐⭐⭐⭐ — The methodology is developed in a clear, layered manner with detailed matrix derivations, though the notation is dense.
Value: ⭐⭐⭐ — Offers practical contributions to the Kansa methods community; the systematic comparison has reference value, but the overall scope of impact is relatively limited.