search

cagrikymk / JAX-ReaxFF

JAX-ReaxFF: A Gradient Based Framework for Extremely Fast Optimization of Reactive Force Fields
GNU General Public License v3.0
53 stars 22 forks source link

About nan values in result #14

Open yjyyyyyyyuuu opened 8 months ago

yjyyyyyyyuuu commented 8 months ago

Dear developers,

I would like to ask you about the reasons why "true loss value" and "loss value after loss minim" have a large number of nan values during the training process. Does this have anything to do with the range of parameters provided or some basic parameter Settings?

Thank you for your time!

xxlliu commented 6 months ago

I have the same question. Have you found the answer? Also look forward to the developer's answer.

Thank you for your time!

cagrikymk commented 4 months ago

Sorry for the delay, I was busy with my final defense for my phd and now I have more time with development.

I have a new branch that should remedy this issue as it is better tested.

https://github.com/cagrikymk/JAX-ReaxFF/tree/jax-reaxff-v0.2

cagrikymk commented 4 months ago

You can rerun the training with the updated code.

However, there is a chance that your starting parameters / training data are not so compatible which might result in high loss value.