Symbolic trial-function method inspired by neural network architectures
Authors:
- Hang-Jing XiongHXHang-Jing Xiong
- Ke-Yan DaiKDKe-Yan Dai
- Jiang-Long ShenJSJiang-Long Shen
- Jing-Bin LiangJLJing-Bin Liang
Open Access
- 1 January 2026
- journal article
- Published by American Institute of Mathematical Sciences (AIMS) in Networks and Heterogeneous Media
- Vol. 21 (5) , 1800-1821
Abstract
This paper proposes a symbolic computational method using neural networks to solve Boussinesq equations. By combining activation functions such as sine, hyperbolic tangent, hyperbolic secant, and exponential, we constructed a single-hidden-layer model [3-3-1] and a double-hidden-layer model [3-3-2-1] to solve the (2+1)-dimensional and (1+1)-dimensional Boussinesq equations, respectively. This yielded multiple sets of novel, exact analytical solutions, which were visualized using 3D plots, contour plots, heatmaps, and evolution plots. To address the issues of insufficient direct neural network feature representation and poor generalization ability, this paper introduces a feature-enhanced neural network method. Model [(2-5)-3-1] was constructed for the (1+1)-dimensional equation, expanding the number of input neurons from 2 to 5; Model [(3-9)-3-1] was constructed for the (2+1)-dimensional equation, expanding the number of input neurons from 3 to 9. Both models retain 3 hidden-layer neurons and optimize the network structure based on the equation's dimension. By reconstructing the physical, spatiotemporal, and nonlinear characteristics of the equations and coupling them with inherent constraints, this method offers the advantages of richer solution results, higher accuracy, and significantly improved model generalization. It does not require bilinear transformations, features a simple structure, and has low computational cost, providing a new, viable solution for solving strongly nonlinear water wave equations.Keywords
This publication has 53 references indexed in Scilit:
- Boussinesq equation solved by the physics-informed neural networksNonlinear Dynamics, 2023
- Periodic-soliton and periodic-type solutions of the (3+1)-dimensional Boiti–Leon–Manna–Pempinelli equation by using BNNMNonlinear Dynamics, 2021
- Numerical solutions of Boussinesq equation using Galerkin finite element methodNumerical Methods for Partial Differential Equations, 2020
- Stability Near Hydrostatic Equilibrium to the 2D Boussinesq Equations Without Thermal DiffusionArchive for Rational Mechanics and Analysis, 2020
- New integrable Boussinesq equations of distinct dimensions with diverse variety of soliton solutionsNonlinear Dynamics, 2019
- Bilinear neural network method to obtain the exact analytical solutions of nonlinear partial differential equations and its application to p-gBKP equationNonlinear Dynamics, 2019
- A numerical method based on extended Raviart–Thomas (ER‐T) mixed finite element method for solving damped Boussinesq equationMathematical Methods in the Applied Sciences, 2017
- Solitary wave solutions of three special types of Boussinesq equationsNonlinear Dynamics, 2017
- Global existence of solutions to the periodic initial value problems for two-dimensional Newton-Boussinesq equationsApplied Mathematics and Mechanics, 2010
- Multi-symplectic method for generalized Boussinesq equationApplied Mathematics and Mechanics, 2008