Application of neural network potentials to modelling transition states

Area of Science:

• Computational Chemistry
• Chemical Dynamics
• Machine Learning in Chemistry

Background:

• Transition state modeling is crucial but computationally expensive, often requiring extensive human input and iterative calculations.
• Exploring the free energy surface (FES) and conformational space around transition states is vital for understanding reaction mechanisms.
• Traditional methods like Density Functional Theory (DFT) can be resource-intensive for thorough exploration.

Purpose of the Study:

• To develop a more efficient computational approach for transition state modeling.
• To reduce the effort and computational cost associated with identifying transition state structures.
• To assess the utility of machine learning potentials for exploring reaction pathways.

Main Methods:

• Utilized umbrella sampling combined with a machine learning potential (ANI-2x) to explore the free energy surface.
• Applied the method to two distinct chemical reactions: amide formation and disulphide bridge formation.
• Compared the efficiency and thoroughness of sampling against traditional DFT methods.

Main Results:

• The machine learning approach demonstrated enhanced efficiency in exploring the free energy surface compared to DFT.
• Rapid and thorough sampling of reaction pathways was achieved, aiding in structure identification.
• ANI-2x showed limitations in accurately predicting high-energy structures but excelled in pathway exploration.

Conclusions:

• The presented method offers a more efficient alternative for exploring conformational space around transition states.
• Machine learning potentials can accelerate the sampling of reaction pathways, guiding further high-level theoretical calculations.
• This approach is valuable for preliminary investigations and hypothesis generation in computational chemistry studies.