Fast uncertainty estimates in deep learning interatomic potentials.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Deep learning models excel at predicting molecular and material properties.
• Current methods lack predictive uncertainty estimates, relying on computationally expensive deep ensembles.
• Quantifying prediction uncertainty is crucial for reliable model application.

Purpose of the Study:

• To develop a computationally efficient method for uncertainty quantification in deep learning.
• To enable accurate uncertainty estimates from a single neural network, avoiding ensemble overhead.
• To validate the proposed method against established deep ensemble techniques.

Main Methods:

• A novel single-neural-network approach for estimating predictive uncertainty.
• Comparison of uncertainty estimates with deep ensembles across configuration space.
• Evaluation of the method's efficacy in active learning scenarios.

Main Results:

• The proposed method achieves uncertainty quantification quality comparable to deep ensembles.
• Uncertainty estimates align with the potential energy surface of the test system.
• The method demonstrates effectiveness in active learning, matching ensemble strategies at reduced cost.

Conclusions:

• A computationally efficient and accurate method for uncertainty quantification using single neural networks has been developed.
• This approach significantly reduces the overhead associated with traditional ensemble methods.
• The findings have implications for accelerating scientific discovery through more efficient machine learning models.