Exploring Neural Network Surrogates for High-Order Mesh-Free Interpolants

Abstract

Mesh-free numerical methods offer flexibility in the discretisation of complex geometries, showing significant potential for problems where mesh-based methods struggle. Although high-order approximations can be obtained through consistency-correction linear systems, such approaches remain prohibitively expensive for Lagrangian formulations, which commonly exhibit low-order convergence. Here we investigate the use of machine learning (ML) to bridge this gap, developing two strategies to couple multilayer perceptrons (MLPs) with the Local Anisotropic Basis Function Method (LABFM) as an exemplar high-order mesh-free method. In the first strategy, neural networks are trained to directly surrogate the high-order kernel; in the second, surrogate models are developed to compute the solution of the dense, low-rank linear systems arising in high-order mesh-free discretisations. The first strategy yields qualitative agreement with validation data but only marginally outperforms inconsistent smoothed particle hydrodynamics (SPH) kernels, with divergent behaviour observed for the Laplacian operator. The second strategy produces solution vectors with mean absolute errors of $\mathcal{O}(10^{-4}$--$10^{-5})$, replicating LABFM second-order convergence rate up to a resolution-dependent limiting accuracy and achieving up to a $5\times$ speedup at equivalent accuracy. However, sensitivity analyses reveal that higher-order approximations impose increasingly stringent accuracy requirements on the predicted solution vector, representing a fundamental challenge for current neural network architectures.