GULP use machine learning interatomic potential

Dear Prof. Julian,

The empirical force field can address the most of issues, but it sometimes has low accuracy.

In addition, it is difficult to obtain the force field parameters for some systems. The machine learning interatomic potential (MLIP) trained from DFT has high accuracy and efficiency. So I wonder if GULP can use the machine learning interatomic potential (MLIP). If not, does GULP consider MLIP in the future as well? Thanks very much.

Best regards

It’s true that machine learning potentials is a good route for some cases, but it should be noted that it’s not always the most efficient route if you factor in the cost of generating the DFT data, but if you have this already then it’s fine. Low accuracy with force fields usually comes from insufficient physics being built into the model and (like machine learning) not enough data going into the fitting.

Now to the question - yes, GULP may directly incorporate MLIPs in future. For now you can already use these for first derivatives via OpenKIM. Just compile and link to OpenKIM and you can use any of the potentials available in this package.

Thanks very much for reply.
Best regards