SAMPLE: Surface structure search enabled by coarse graining and statistical learning

Studying the electronic structure of organic monolayers on inorganic substrates requires knowledge about their atomistic structure. Such monolayers often display rich polymorphism arising from diverse molecular arrangements in different unit cells. The large number of possible arrangements poses a considerable challenge for determining the different polymorphs from first principles.

To meet this challenge, we developed SAMPLE[1-3], which employs coarse-grained modeling and machine learning to efficiently map the minima of the potential energy surface of commensurate organic adlayers. Requiring only a few hundred DFT calculations of possible polymorphs as input, we use Bayesian linear regression to determine the parameters of a physically motivated energy model. These parameters yield meaningful physical insight and allow predicting adsorption energies for millions of possible polymorphs with high accuracy.

We demonstrate SAMPLEs capabilities on the systems of naphthalene[1] and TCNE[2,3] on coinage metals where we predict the energetically most favorable polymorphs and compare them to experimental data.

[1] Hörmann et al., Comput. Phys. Commun. 244, 143–155, 2019
[2] Scherbela et al., Phys. Rev. Materials 2, 043803, 2018
[3] Obersteiner et al., Nano Lett. 17, 4453-4460, 2017