Article contents
Maximum-Principle-Certified Adaptive IMEX–Neural Solver for Stiff Nonlinear Reaction–Diffusion Equations with Heterogeneous Diffusion
Abstract
High reaction rates, along with heterogeneous diffusion, cause reaction–diffusion coupling to be time step-sensitive, and uncontrolled neural correction can place an otherwise stable point out of the physical domain. In this way, a hybrid method is constructed on top of a finite-volume stabilized IMEX predictor. At each accepted time step, a radial basis compact neural network is trained using the backward Euler residual rather than labelled solution samples. Inward limiting and a dyadic acceptance test restrict the correction to occur only if the state stays inside and the residual norm doesn’t increase. Time step doubling ensures the temporal accuracy independent of the neural training. Finite volume diffusion matrix leads to an -matrix proof of the discrete maximum principle, and fallback ensures certificate availability for each time step. The fixed-step relative error for the 1D heterogeneous Fisher-KPP problem is reduced from to . For the 1D Nagumo front problem, the error reduction is from to ; however, the crude error correction exceeds the bound by . The bound violation by the verified trajectory is zero. The 2D heterogeneous Fisher-KPP test validates the residual reduction and adaptive gain but with an expense incurred by the local error correction.
Article information
Journal
British Journal of Multidisciplinary Studies
Volume (Issue)
4 (2)
Pages
76-86
Published
Copyright
Copyright (c) 2026 https://creativecommons.org/licenses/by/4.0/
Open access

This work is licensed under a Creative Commons Attribution 4.0 International License.
Aims & scope
Call for Papers
Article Processing Charges
Publications Ethics
Google Scholar Citations
Recruitment