Implicit neural representations for chemical reaction paths.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Traditional methods like Nudged Elastic Band (NEB) discretely approximate minimum energy paths.
• These methods can struggle with complex systems, poor initial guesses, or multiple competing reaction pathways.

Purpose of the Study:

• To develop a continuous representation of minimum energy paths using neural networks.
• To offer a flexible and efficient alternative to discrete path-search algorithms.
• To demonstrate the method's applicability to challenging atomistic systems.

Main Methods:

• Parameterizing reaction paths with a neural network.
• Training the network using a loss function that discards tangential energy gradients.
• Validating the approach on 2D potentials and complex atomistic systems.

Main Results:

• The neural network method provides a continuous function for minimum energy paths.
• It successfully estimates transition states and outperforms NEB on challenging systems.
• Adjusting sampling strategies aids in escaping local minima.
• The network demonstrated generalization to unseen systems.

Conclusions:

• Neural networks offer a versatile and continuous approach to representing minimum energy paths.
• This method enhances the study of chemical reactions and material transformations.
• The potential for a universal reaction path representation using machine learning is promising.