Can NequIP use ZBL to prevent atoms from being too close together?

[ChengQIAN93]

Hi, everyone,

I've been using NequIP to train my system. I added structures with short interatomic distances (e.g., 1.0 to 1.5 angstroms) to the training set to prevent atoms from being too close together.

However, some simulations show that the interatomic distance is still too close, which makes a very low potential energy.

My quesiton is can NequIP use ZBL or another method to make atoms repel each other when they're too close?

Thank you.

[cw-tan]

Hi @ChengQIAN93 ,

Thank you for your interest in NequIP. It is possible to include the ZBL term in a NequIP GNN model as shown in this tutorial config file. The nuclear repulsion effect of ZBL should provide a strong prior to make atoms repel each other when they get too close, but it could be possible that either the repulsion is too short-range (it would also depend on the polynomial cutoff's p hyperparameter as that dictates how the ZBL term decays to zero at the radial cutoff), or that the NequIP model was initialized in a way that counteracts the ZBL near atom cores (though this effect is unlikely to be too significant). We have seen empirically that ZBL helps make MD more stable by preventing atoms from getting too close to each other, so it's definitely worth a shot. Adding structures with short interatomic distances would be helpful too, and a combination of both ZBL and including such frames would likely be useful.

[ChengQIAN93]

Thanks for your kindly reply. I tried your suggestion of using ‘pair_style’ and it worked well. I have one more question. If I need to use ‘zbl’, do I need to include some structures with short distances for the fitting of zbl parameters?

[cw-tan]

There's no fitting of ZBL parameters, the parameters are deterministic from the atomic species. See code here. In general though, it would be good to include some structures with short distances to condition the model on them. While the ZBL term is deterministic, I don't have a good sense on how the untrained NequIP GNN model will behave for the short distance regime and how it would evolve during training if it's never seen atomic environments with short atomic distances -- adding those structures to your training should make the overall model more robust (learned NequIP GNN model + ZBL potential).