Nequip 0.7.0 with energy labels only

[Mehryar-jn]

Hello,

In previous versions (e.g., v0.6.x), there were example YAML configs for training on datasets with only energy labels (no forces), such as configs/minimal_eng.yaml. I couldn't find a similar example for v0.7.0, and the config format seems to have changed quite a bit. Also, I noticed that the user guide doesn't provide direct instructions for this use case. Could you please share an updated minimal config for this scenario?

Thanks!

[cw-tan]

You shouldn't have to change anything in the model -- it'll still predict forces, etc, but you just won't train on them (if that's a painful overhead, we can also do something about that). All you need is to use loss and metrics that only include energy terms.

I would direct you to:

• https://nequip.readthedocs.io/en/latest/guide/configuration/metrics.html, and
• https://nequip.readthedocs.io/en/latest/api/metrics.html#advanced-configuration-metricsmanager

In particular, you might find the config file chunk from the docstring of MetricsManager to be helpful. For loss functions, it is typical to use MSEs. For metrics (val, test, etc), it is typical to monitor MAEs and RMSEs. We only provide simplified wrappers for conventional force field training scenarios, i.e. energy and force (or energy, force, and stress). So you have to manually configure the base MetricsManager with only force metrics.

We are open to having more energy-only support if there is a good reason for it.

[Mehryar-jn]

Thanks for your prompt reply!

I'm using the MetricsManager for loss and metrics as below:

loss:
target: nequip.train.MetricsManager
metrics:
- name: per_atom_energy_mse
field:
target: nequip.data.PerAtomModifier
field: total_energy
coeff: 1
metric:
target: nequip.train.MeanSquaredError

val_metrics:
target: nequip.train.MetricsManager
metrics:
- name: per_atom_energy_rmse
field:
target: nequip.data.PerAtomModifier
field: total_energy
coeff: 1
metric:
target: nequip.train.RootMeanSquaredError

However, I get the following error:
KeyError: 'forces'

Do I need to add dummy forces to my dataset or is the model trying to access forces in some other parts like dataset statistics?

[cw-tan]

ah, sorry I missed dataset statistics: https://nequip.readthedocs.io/en/latest/guide/configuration/data.html#dataset-statistics

You can do something similar to configure the base stats manager class: https://nequip.readthedocs.io/en/latest/api/data_stats.html#nequip.data.DataStatisticsManager

You can use the standard deviation of the total energy or per-atom energy as the scales. If you have access to isolated atom energies it would also be important to use them for the shifts. See https://nequip.readthedocs.io/en/latest/guide/configuration/model.html#training-data-statistics-as-hyperparameters

[Mehryar-jn]

Thanks a lot! DataStatisticsManager fixed it!

[cw-tan]

Hi @Mehryar-jn , as of NequIP v0.11.0, we provide specific energy-only utilities that you can use to reduce the amount of boilerplate in your config: https://nequip.readthedocs.io/en/latest/guide/reference/faq.html#energy-only-training

[Mehryar-jn]

Hi @cw-tan , that's great! Thanks.