Neural network approach to time-dependent dividing surfaces in classical reaction dynamics

Area of Science:

• Chemical Dynamics
• Computational Chemistry
• Machine Learning Applications

Background:

• Chemical reaction rates are governed by energy barriers and the flux through a dividing surface (DS) between reactants and products.
• Accurate rate calculations require DSs free of recrossings, which are challenging to construct, especially for time-dependent or high-dimensional systems.
• Traditional methods like transition state theory face computational demands in determining precise DS geometries.

Purpose of the Study:

• To introduce a novel method for constructing time-dependent, global, and recrossing-free dividing surfaces (DS) using neural networks.
• To enable accurate determination of reaction rates in dynamical systems, overcoming limitations of existing computational approaches.
• To provide a computationally efficient and generalizable technique for phase space separation in chemical dynamics.

Main Methods:

• Utilized neural networks that take bath coordinates and time as input to predict the DS position along the reaction coordinate.
• Trained the neural network to ensure the constructed DS is free of recrossings, crucial for accurate rate calculations.
• Applied the method to two- and three-dimensional dynamical systems to demonstrate its efficacy and generalizability.

Main Results:

• Successfully constructed time-dependent, global, and recrossing-free DSs using neural networks.
• Demonstrated that trained neural networks store complete phase space separation information, allowing precise reactant-product distinction.
• Showcased the method's applicability in 2D and 3D systems with potential for extension to higher dimensions.

Conclusions:

• Neural networks offer a powerful and computationally efficient tool for constructing accurate dividing surfaces in dynamical systems.
• This approach significantly simplifies the determination of reaction rates by providing a robust method for phase space partitioning.
• The developed technique has broad implications for computational chemistry and chemical dynamics, particularly for complex, high-dimensional systems.