Bruno Sixou

Vikas Dwivedi, Balaji Srinivasan, Monica Sigovan et al.

This paper introduces the Kernel Adaptive Physics-Informed Extreme Learning Machine (KAPI-ELM), an adaptive Radial Basis Function (RBF)-based extension of PI-ELM designed to solve both forward and inverse Partial Differential Equation (PDE) problems involving localized sharp gradients. While PI-ELMs outperform the traditional Physics-Informed Neural Networks (PINNs) in speed due to their single-shot, least square optimization, this advantage comes at a cost: their fixed, randomly initialized input layer limits their ability to capture sharp gradients. To overcome this limitation, we introduce a lightweight Bayesian Optimization (BO) framework that, instead of adjusting each input layer parameter individually as in traditional backpropagation, learns a small set of hyperparameters defining the statistical distribution from which the input weights are drawn. This novel distributional optimization strategy -- combining BO for input layer distributional parameters with least-squares optimization for output layer network parameters -- enables KAPI-ELM to preserve PI-ELM's speed while matching or exceeding the expressiveness of PINNs. We validate the proposed methodology on several challenging forward and inverse PDE benchmarks, including a 1D singularly perturbed convection-diffusion equation, a 2D Poisson equation with sharp localized sources, and a time-dependent advection equation. Notably, KAPI-ELM achieves state-of-the-art accuracy in both forward and inverse settings. In stiff PDE regimes, it matches or even outperforms advanced methods such as the Extended Theory of Functional Connections (XTFC), while requiring nearly an order of magnitude fewer tunable parameters. These results establish the potential of KAPI-ELM as a scalable, interpretable, and generalizable physics-informed learning framework, especially in stiff PDE regimes.

Vikas Dwivedi, Enrico Schiassi, Monica Sigovan et al.

Physics-informed neural networks (PINNs) and related methods struggle to resolve sharp gradients in singularly perturbed boundary value problems without resorting to some form of domain decomposition, which often introduce complex interface penalties. While the Extreme Theory of Functional Connections (X-TFC) avoids multi-objective optimization by employing exact boundary condition enforcement, it remains computationally inefficient for boundary layers and incompatible with decomposition. We propose Gated X-TFC, a novel framework for both forward and inverse problems, that overcomes these limitations through a soft, learned domain decomposition. Our method replaces hard interfaces with a differentiable logistic gate that dynamically adapts radial basis function (RBF) kernel widths across the domain, eliminating the need for interface penalties. This approach yields not only superior accuracy but also dramatic improvements in computational efficiency: on a benchmark one dimensional (1D) convection-diffusion, Gated X-TFC achieves an order-of-magnitude lower error than standard X-TFC while using 80 percent fewer collocation points and reducing training time by 66 percent. In addition, we introduce an operator-conditioned meta-learning layer that learns a probabilistic mapping from PDE parameters to optimal gate configurations, enabling fast, uncertainty-aware warm-starting for new problem instances. We further demonstrate scalability to multiple subdomains and higher dimensions by solving a twin boundary-layer equation and a 2D Poisson problem with a sharp Gaussian source. Overall, Gated X-TFC delivers a simple alternative alternative to PINNs that is both accurate and computationally efficient for challenging boundar-layer regimes. Future work will focus on nonlinear problems.