What is BOSS

BOSS uses Bayesian optimization (BO) active learning method to iteratively build a surrogate model of the target property (for example potential energy of a system). BO employs uncertainty-led exploration/exploitation sampling strategy, which allows quick convergence of the model with a modest number of data acquisitions. The target property is optimized via automatic detection of global and local minima, which are reliably extracted from the N-dimensional surrogate model. The model is typically constructed in a phase space of under 10 dimensions. This is achieved for example in atomistic systems by describing the structure with chemical building blocks (see Example 1).

The main features of BOSS include automatic Gaussian process (GP) hyperparameter optimization for reliable model convergence, various acquisition functions for different BO tasks, and a wide range of postprocessing tools to extract information from the model. With a simple user function Python script, BOSS can be interfaced to many different kinds of simulation software and experimental procedures.

Capabilities at a glance

The following sections outline the basic features of BOSS, such as building the model with active learning, data-mining the model via postprocessing, and other practical features.

Further information

Further information is provided at the BOSS main website with various different research examples. The BOSS code is available for download in GitLab.

Authors

BOSS was conceived by Milica Todorović (University of Turku) and Patrick Rinke (Aalto University) in collaboration with Jukka Corander (University of Helsinki/University of Oslo) and Michael Gutmann (University of Edinburgh). The original BOSS prototype MATLAB code was developed by MT and MG.

The BOSS python package is under continuous development at Aalto University by the Computational Electronic Structure Theory (CEST) group and at University of Turku.

See BOSS people for the full list of authors.