Comparing interactive molecular dynamics and constrained molecular dynamics

The Glowacki group recently published a new paper in the Journal of Physical Chemistry A about fitting Potential Energy Surfaces.

D1D2
Hydrogen abstraction by a Cyano radical

We compared surfaces of the reaction of Isopentane with a Cyano radical obtained from fitting two different data sets: 1) sampled with interactive molecular dynamics in Virtual Reality and 2) sampled with constrained molecular dynamics.

KDE
A) Constrained molecular dynamics data B) Interactive Molecular Dynamics data

The aim was to see whether one surface is considerably better than the other.

To do the fitting, we used an atomic neural network with atom centred symmetry functions as the representation for the atoms. We fitted a network to each data set separately, i.e. – we fitted one network to the data set obtained with interactive molecular dynamicsand one network to the data set obtained with constrained molecular dynamics.

blog1
Training 2 separate networks on two different data sets.

Then, we used each neural network to predict the energies of the structures from the other data set, i.e. – we used the network trained on the interactive molecular dynamics data set to predict the energies of the structures in the constrained molecular dynamics data set and vice versa.

blog2
Using the previously trained networks to predict the energies of the opposite data sets.

This showed that the neural network trained on the interactive molecular dynamics data set is better at predicting the energies of more stable structures, while the network trained on the constrained molecular dynamics data set does better for the high energy structures. This can be seen from the correlation plots below, that shows the true constrained molecular dynamics (CMD) and interactive molecular dynamics (iMD-VR) energies compared with the predictions.

cross_pred
Comparison of A) the true constrained MD energies and their prediction by the network trained on the interactive MD data and B) vice versa.

We also tried to predict the energies of an optimised reactions surface for a primary hydrogen abstraction with both networks. Neither the interactive molecular dynamics or the constrained molecular dynamics data set contained optimised structures. In this case, the network trained on the constrained molecular dynamics data set did better.

surface_comparison.png
Predicted reaction surface for a primary hydrogen abstraction by A) the network trained on the interactive MD data and B) the network trained on the constrained MD data.

We concluded that the interactive molecular dynamics method is a viable alternative to constrained molecular dynamics for sampling reactions. It offers the advantage of ease of use once it is set up and enables to easily sample close to the minimum energy path.

One thought on “Comparing interactive molecular dynamics and constrained molecular dynamics

Add yours

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

Blog at WordPress.com.

Up ↑

%d bloggers like this: