Lately, I've become quite intrigued by the recent advances in machine learning forcefields within the field of computational chemistry. I've developed a curiosity about training a simple forcefield on a transition metal bulk mono-element.

I've been pondering the best approach to train the model. Initially, the idea that came to mind was training it on different volumes, elementary transformations (such as shear and strain), and normal modes (and combinations of these three).

To gain more insights, I've been reading the literature, and it appears that various approaches are employed:

  • Training on completely random structures: There are different methods available, but the underlying result is that you predominantly train on structures that may not have experimental relevance and, in some cases, aren't physically plausible. One potential advantage is that your forcefield could become robust against diverse scenarios. However, the way I understand it, there is a possible accuracy tradeoff on important properties for real systems?

  • Training on Molecular Dynamics data of known systems (FCC here): This particular approach puzzles me a bit, as it seems to be the opposite of the previous one. Essentially, you train the model on correlated and biased data, often at a specific temperature. At first glance, this doesn't appear to be advantageous over randomly generated displacements (uncorrelated)?

At the moment, my training approach involves scanning different volumes and elementary transformations, while incorporating random displacements of varying intensities.

Obviously, how to train is heavily system and potential dependent. I understand that this is quite a vague question. I would be grateful if someone could give me an insight, or (even better) correct me if I wrote anything wrong.


1 Answer 1


I'll do my best to answer your question because I'm also currently learning about this topic. The field generated by machine learning is indeed receiving more and more attention from everyone. However, in my opinion, to obtain a more accurate field, different machine learning methods and models require rich datasets, and of course, the larger the dataset, the better.

Regarding your question, theoretically, the more extensive the training configuration library, the better the predictive power for unknown structures in the production run. Suppose you are studying the phase transition process of a crystal structure; in that case, in addition to the structures of these different phases in your training library, the transitional structures between them also need good training. Therefore, what I want to say is that the most crucial aspect of a machine learning training field method is not how many configuration libraries it can obtain, but how clearly it can distinguish which configuration libraries are worth learning. For example, if you have a dataset of 10,000 AIMD data, only 300 of them may be necessary for training. Therefore, for your vague question, my answer is to try to understand the process you want to simulate as much as possible (using AIMD to simulate small models is a good method). Additionally, I recommend that you first use some ready-made machine learning methods and then improve upon them. For example, VASP's MLFF function provides an on-the-fly learning mode, which can reduce the training time of traditional machine learning. However, its shortcomings are also evident, as the kernel model used has limited adjustability and is unsuitable for studying many structures. Secondly, there is the DeepMD kit developed by E. Weinan, which is also a relatively mature kit for generating machine learning fields. You can start with these kits for training and then improve upon them. I believe that this can give you a deeper understanding of your own machine learning and the problems you want to study.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .