Deep ensembles vs committees for uncertainty estimation in neural-network force fields: Comparison and application to...

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Reliable uncertainty estimation is crucial for machine-learning force fields.
• Current methods like simple committees for neural-network force fields have limitations.

Purpose of the Study:

• To present a generalized deep-ensemble design for uncertainty estimation in neural-network force fields.
• To incorporate aleatoric uncertainty from training data.
• To enable efficient force field refinement through active learning.

Main Methods:

• A generalized deep-ensemble design using multiheaded neural networks and a heteroscedastic loss.
• Comparison of uncertainty metrics from deep ensembles, committees, and bootstrap-aggregation ensembles.
• An adversarial active learning approach for progressive force field refinement.

Main Results:

• The proposed method efficiently estimates uncertainties in both energy and forces.
• Demonstrated effectiveness on ionic liquid and perovskite surface data.
• Active learning workflow is feasible due to fast training with residual learning and a nonlinear learned optimizer.

Conclusions:

• The generalized deep-ensemble method offers a robust and efficient way to estimate uncertainties in machine-learning force fields.
• This work facilitates the development of more reliable and accurate predictive models.
• The proposed active learning strategy accelerates force field improvement.