search

mir-group / nequip

NequIP is a code for building E(3)-equivariant interatomic potentials
https://www.nature.com/articles/s41467-022-29939-5
MIT License
565 stars 124 forks source link

MLFF for Silicon #407

Open knc6 opened 5 months ago

knc6 commented 5 months ago

Training a silicon MLFF with nequip shows reasonable force error but high energy MAEs. A colab notebook is kept here . Any thoughts on what could be going wrong in terms of setting config parameters etc.?

cw-tan commented 2 weeks ago

Hi @knc6,

Thank you for your interest in using NequIP and apologies for the delay! Maybe you've figured it out by now, but it appears that the config file you're using only includes forces in the loss function, i.e. 

# loss function
loss_coeffs: forces

You can also include energies in the loss function as follows

# loss function
loss_coeffs:                                                                        
  forces: 1
  total_energy:                                                                    
    - 1
    - PerAtomMSELoss

For example, see this example.yaml.