Intelligent Time-Scale Operator-Splitting Integration for Chemical Reaction Systems.

Area of Science:

• Computational Chemistry
• Chemical Engineering
• Numerical Analysis

Background:

• Numerical stiffness in reactive flow simulations arises from wide ranges of time scales in chemical reactions.
• Existing methods for handling stiffness, like eigendecomposition, are computationally expensive.

Purpose of the Study:

• To develop an intelligent time-scale operator-splitting (OS) chemistry integration method.
• To reduce numerical stiffness and model complexity in reactive flow simulations.
• To replace computationally intensive Jacobian eigendecomposition with a neural network approach.

Main Methods:

• An operator-splitting (OS) method is proposed for chemistry integration.
• A pretrained backpropagation neural network identifies slow/fast reactions and stiffness sources.
• Fast-slow decomposition separates the chemical source term into stiff and non-stiff components.
• Stable time-scale OS integration solves stiff chemical ordinary differential equations.

Main Results:

• The intelligent OS method effectively reduces numerical stiffness and model complexity.
• The neural network approach avoids expensive eigendecomposition of the Jacobian matrix.
• The method balances computational cost with accuracy in simulations.
• Favorable comparisons were made against implicit Euler, explicit Euler, and Runge-Kutta solvers.

Conclusions:

• The proposed intelligent time-scale OS method is effective for stiff chemical ordinary differential equations in reactive flow.
• This approach offers a computationally efficient and accurate alternative to traditional solvers.
• The use of neural networks provides an on-the-fly identification of reaction dynamics and stiffness.