Benchmarking electronic structure models is an important performance assessment effort for molecular science. Traditionally, a benchmark set is defined first and only then a range of models is benchmarked against reference data. However, it is questionable how the corresponding performance measures that are usually based on average-error statistics can be transferred to molecular systems outside the benchmark set, in particular if their dissimilarity to the benchmark systems is rather large. To resolve this issue, we employ a Bayesian decision maker that, given a user query for new molecular systems, provides yes/no-type recommendations on whether a new reference value is necessary or not. As the decision maker learns from previous reference data, we can steadily reduce the number of reference measurements/calculations required to achieve a predefined accuracy. An appealing consequence of our Bayesian approach are system-specific (instead of average) performance measures, which allow for the assessment of electronic structure models on a case-by-case basis. The decisions to be made are based on machine-learned variances that are inverse measures of how certain the decision maker is about the accuracy of the estimated system-specific performances. Our approach is particularly appealing for high-throughput studies where the extent of queries can no longer be managed by a human expert. We demonstrate the workflow for a search space comprising ~10<sup>6</sup> synthetically accessible molecular organic semiconductors and benchmark a diverse set of DFT-based and semiempirical models for excited-state properties that are of general interest to the molecular electronics community.

Automated generation of benchmark sets guided by a Bayesian decision maker

2019 Fall National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.