Cluster expansions have proven to be a valuable tool in alloy theory and other problems in materials science but the generation of cluster expansions can be a computationally expensive and time-consuming process. We present a Bayesian framework for developing cluster expansions that explicitly incorporates physical insight into the fitting procedure. We demonstrate how existing methods fit within this framework and use the framework to develop methods that significantly improve the predictive power of cluster expansions for a given training set size. The key to the methods is to apply physical insight and cross validation to develop physically meaningful prior probability distributions for the cluster expansion coefficients. We use the Bayesian approach to develop an efficient method for generating cluster expansions for low-symmetry systems such as surfaces and nanoparticles.

• Received 30 December 2008

DOI:https://doi.org/10.1103/PhysRevB.80.024103