A multilevel stochastic-gradient neural solver for boundary integral equations
Abstract
We propose a multilevel stochastic-gradient neural solver (MLSG) for second-kind boundary integral equations. MLSG represents the unknown boundary density using a neural network optimized by stochastic residual minimization over a hierarchy of successively refined Nyström discretizations. Upon transitioning from one level to the next, the network parameters obtained on the previous level initialize training on the current one. This coarse-to-fine strategy retains a continuous, grid-independent density representation and is designed to reduce the total computational effort required to reach a prescribed residual tolerance at the target resolution. The algorithm avoids grid-transfer operators and hierarchical fast-summation machinery, relying instead on batched kernel evaluations and standard network forward and backward passes that map efficiently onto modern GPU architectures. For uniformly stable second-kind discretizations, so strongly nonuniform contraction rates originate in the empirical neural tangent kernel (NTK) rather than in the discretized operator. Within each level, parameter updates can reshape the NTK, while refinement re-samples the tangent kernel on a richer discrete space and reveals directions that were not adequately resolved on coarser levels. A cross-level estimate bounds the warm-start loss in terms of the preceding training tolerance and quadrature error, motivating a tolerance schedule that balances optimization and discretization errors. Experiments on Laplace/Poisson and Helmholtz problems in two and three dimensions, together with an exterior Robin problem on a hypersurface in R4, demonstrate the method under both parametric and signed-distance surface representations at up to million-scale discretizations.
Versions
Related
- Multilevel Picard approximations for McKean-Vlasov stochastic differential equations with jumps
- Existence and multiplicity results for third-order ordinary differential equations with derivative dependence subject to the Riemann–Stieltjes integral boundary conditions
- Singularity Analysis of Laplace Boundary Integral Equations on Three-Dimensional Domains with Corners
- Hierarchical low-rank compression exploiting symmetric boundary integral equations in transmission scattering
- Kernel-free Boundary Integral Methods for Allen-Cahn and Cahn-Hilliard Equations on Irregular Domains
BibTeX
@misc{lu2026multilevel,
title = {{A multilevel stochastic-gradient neural solver for boundary integral equations}},
author = {Bing-Ze Lu and Richard Tsai},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2607.00560},
eprint = {2607.00560},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2607.00560},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - A multilevel stochastic-gradient neural solver for boundary integral equations
AU - Bing-Ze Lu
AU - Richard Tsai
PY - 2026
DA - 2026/10/19
PB - arXiv
DO - 10.48550/arxiv.2607.00560
UR - https://arxiv.org/abs/2607.00560
AB - We propose a multilevel stochastic-gradient neural solver (MLSG) for second-kind boundary integral equations. MLSG represents the unknown boundary density using a neural network optimized by stochastic residual minimization over a hierarchy of successively refined Nyström discretizations. Upon transitioning from one level to the next, the network parameters obtained on the previous level initialize training on the current one. This coarse-to-fine strategy retains a continuous, grid-independent density representation and is designed to reduce the total computational effort required to reach a prescribed residual tolerance at the target resolution. The algorithm avoids grid-transfer operators and hierarchical fast-summation machinery, relying instead on batched kernel evaluations and standard network forward and backward passes that map efficiently onto modern GPU architectures. For uniformly stable second-kind discretizations, so strongly nonuniform contraction rates originate in the empirical neural tangent kernel (NTK) rather than in the discretized operator. Within each level, parameter updates can reshape the NTK, while refinement re-samples the tangent kernel on a richer discrete space and reveals directions that were not adequately resolved on coarser levels. A cross-level estimate bounds the warm-start loss in terms of the preceding training tolerance and quadrature error, motivating a tolerance schedule that balances optimization and discretization errors. Experiments on Laplace/Poisson and Helmholtz problems in two and three dimensions, together with an exterior Robin problem on a hypersurface in R^4, demonstrate the method under both parametric and signed-distance surface representations at up to million-scale discretizations.
ER -