Loss, Metrics, and Uncertainty

Training objectives and validation scores are configured separately but should use consistent unit weighting.

Loss types

Registered in enerzyme/models/loss.py:

• rmse / mse — root mean square error per target

• mae — mean absolute error

• crps — continuous ranked probability score (useful with shallow ensembles)

• nll — negative log-likelihood where supported

• Architecture-specific terms (e.g. PhysNet nh_penalty)

Example:

loss:
    crps:
        E: 1
    rmse:
        Fa: 0.91656181
        M2: 0.32731016
        Q: 0.1

Metrics and early stopping

Per-model Metric blocks define the validation judge_score — a weighted sum of metric terms. Early stopping uses Trainer.patience on this score.

Metric:
    E:
        rmse: 1
    Fa:
        rmse: 0.91656181

Global Metrics in older configs may still appear; per-model Metric in Modelhub is preferred in current train.yaml.

Shallow ensemble uncertainty

ShallowEnsembleReduce with shallow_ensemble_size > 1 in Core trains multiple forward passes. Configure variance outputs:

- name: ShallowEnsembleReduce
  params:
      reduce_mean: [E, Fa, Qa, M2, Q]
      var: [E, Fa]
      eval_only: true

At predict time, request non-target features:

Trainer:
    non_target_features:
        - E_var
        - Fa_var
        - Qa_var

Committee training

Set Trainer.committee_size > 1 to train independent models (FF_committee). Uncertainty comes from disagreement across committee members rather than a single shallow ensemble. Used heavily in active_learning_train.yaml.

UDD at simulation time

Hybrid simulation can apply uncertainty-driven debiasing via uncertainty_calculator: UDD in Simulation (see uma.yaml and Enhanced Sampling and PLUMED). This is separate from training-time E_var / Fa_var.

Unit weights

See Units and Standard Fields. Mismatched force weights are a common cause of poor force accuracy despite good energy loss.