Bayesian ensemble approach to error estimation of interatomic potentials

Area of Science:

• Computational materials science
• Statistical modeling
• Bayesian inference

Background:

• Accurate prediction of material properties relies on reliable models.
• Quantifying uncertainty in model predictions is crucial for assessing reliability.
• Interatomic potentials are essential for simulating material behavior.

Purpose of the Study:

• Develop a general Bayesian method to estimate error bars on model predictions.
• Apply this method to assess uncertainties in interatomic potentials for molybdenum.
• Evaluate the accuracy of the estimated error bars.

Main Methods:

• Utilized a Bayesian approach to model fitting and uncertainty quantification.
• Generated ensembles of models by sampling parameter space based on minimum cost.
• Applied the method to interatomic potentials for molybdenum using atomic force data.

Main Results:

• The developed method provides robust error bar estimations for model predictions.
• Calculated error bars for elastic constants, surface energies, structural energies, and dislocation properties of molybdenum.
• Demonstrated that the estimated error bars realistically reflect actual potential errors.

Conclusions:

• The Bayesian method offers a reliable way to assess prediction uncertainties in computational materials science.
• Accurate error bars enhance the trustworthiness of interatomic potentials and material simulations.
• This approach is broadly applicable to various modeling tasks in science and engineering.