Local Bayesian Dirichlet mixing of imperfect models.

Area of Science:

• Computational modeling
• Statistical machine learning
• Nuclear physics

Background:

• Complex computational models often face challenges in predicting outcomes in experimentally unknown domains.
• Combining results from multiple imperfect models is a strategy to enhance predictive power.
• Bayesian stacking is a known technique for model combination.

Purpose of the Study:

• To propose a novel Bayesian statistical machine learning framework for improving the predictability of complex computational models.
• To extend existing Bayesian stacking methods using the Dirichlet distribution.
• To evaluate the effectiveness of Bayesian model averaging and mixing techniques for nuclear mass prediction.

Main Methods:

• Development of a Bayesian statistical machine learning framework utilizing the Dirichlet distribution.
• Application of Bayesian model averaging and mixing techniques.
• Analysis of global and local mixtures of models for mining nuclear masses.
• Comparison of mixing techniques against classical Bayesian model averaging.

Main Results:

• Bayesian model averaging and mixing techniques demonstrated excellent performance in prediction accuracy for nuclear masses.
• Both global and local mixtures of models provided superior uncertainty quantification compared to classical Bayesian model averaging.
• The proposed framework effectively combines results from several imperfect models.

Conclusions:

• Global and local mixtures of models are preferable to classical Bayesian model averaging for nuclear mass prediction.
• Improving model predictions through mixing, rather than mixing of corrected models, leads to more robust extrapolations.
• The Bayesian framework offers enhanced predictability and uncertainty quantification in complex computational domains.