A constrained backpropagation approach for the adaptive solution of partial differential equations

Area of Science:

• Computational Mathematics
• Numerical Analysis
• Machine Learning Applications

Background:

• Solving nonlinear partial differential equations (PDEs) adaptively is crucial for modeling complex systems.
• Existing methods often rely on penalty functions or Lagrange multipliers, which can be computationally intensive or introduce inaccuracies.
• Neural networks offer a promising approach for approximating PDE solutions, but effectively handling constraints remains a challenge.

Purpose of the Study:

• To present a novel constrained backpropagation (CPROP) methodology for adaptive solutions of nonlinear elliptic and parabolic PDEs.
• To demonstrate CPROP's ability to handle changing PDE parameters and external forcing.
• To offer an alternative to existing constraint-handling techniques in neural network-based PDE solvers.

Main Methods:

• Developed a constrained backpropagation (CPROP) algorithm for training neural networks to solve PDEs.
• Employed direct elimination to solve the constrained optimization problem, reducing dimensionality.
• Ensured exact satisfaction of boundary and initial conditions at each iteration.

Main Results:

• CPROP effectively solves nonlinear elliptic and parabolic PDEs adaptively.
• The method accurately handles changing PDE parameters and nonhomogeneous terms.
• Demonstrated exact enforcement of boundary and initial conditions throughout the iterative process.

Conclusions:

• Constrained backpropagation (CPROP) provides an efficient and accurate method for solving nonlinear PDEs.
• The direct elimination approach simplifies the optimization problem and guarantees constraint satisfaction.
• CPROP offers a robust framework for adaptive PDE solutions in dynamic scenarios.