My Community

I am investigating the recrystallization dynamics of the 2D material MoS2 from the amorphous phase. I have carried out a first-principles simulation and have observed the details of the crystallization process. I would like to scale up the process using machine learning.

The melting point of the system is quite high (2000K). In my first-principles simulation, after equilibrating at 1800K, I then quenched the system to room temperature. I then heated the system to 1800K and after about a couple of 100 ps, the system crystallized.

I would like to carry out the same recrystallization process with more atoms.. As the computational costs are quite high, I would like to do this using machine learning. I am curious how to optimize the machine learning process. I assume I should start out with the system in the crystalline state and then gradually raise the temperature to just below the melting point with the usual ICONST and NPT boundary conditions. What is a good target for the number of configurations for each atom type? If there is there any additional advice what can be offered, it would be most gratefully received.

First you need the training data for the right phases.
You need to train on the liquid phase and then train on the solid. Possibly in a heating run going up to temperatures just below the melting point.

If you are lucky and the melting of the system occurs during a long MD then I would just simply use heating and cooling runs in the production runs.
Just try it out with a long MD.
If the phase transition won't occur spontanously in a heating or cooling run then you need to use interface pinning.
We have used that method in our MLFF implementation paper.
For that method one needs to construct liquid and solid interfaces. The nice thing in that method is that we found out that one does not have to train on the big interface, but it was enough to train on the liquid and solid phases separately.