Catalyzing Discovery: Bayesian Back End (BayBE) as a general-purpose toolbox for Bayesian Design of Experiments in real-world campaigns

J. Hoche, M. Müller, M. Fitzner, A. Hopp, A. Lee, A. Šošić and J. G. Brandenburg

The Bayesian Back End (BayBE) is a versatile, open-source software package [1] designed to streamline and optimize the design of experiments (DoE) process. BayBE offers a comprehensive suite of tools tailored for real-world experimental campaigns, particularly in the chemical and materials sciences. Key features include custom parameter encodings to incorporate domain knowledge, built-in chemical encodings, and the flexibility to integrate custom surrogate models for specialized problems like active learning.

A standout capability of BayBE is its implementation of transfer learning, enabling the integration of data from multiple related campaigns to accelerate optimization. This multi- context Bayesian optimization approach has demonstrated significant performance enhancements in various test cases, including reaction conditions for direct arylation.[2] and anti-correlated benchmark functions. Preliminary results suggest that strategically sub- sampling source data can further improve optimization efficiency.

BayBE is built on a robust, fully typed and hypothesis-tested code base, ensuring reliability and ease of use. It includes comprehensive backtesting utilities to facilitate benchmarking and identification of optimal settings. All objects within BayBE are fully de-/serializable, making it easy to store results in databases or integrate with wrappers such as APIs.

Figure 1. Outcomes of Bayesian Optimization with various molecular encodings from BayBE [1], demonstrating enhanced efficiency in identifying optima˚l conditions for imidazole direct arylation [2] with complex encodings over simpler ones with fewer iterations.

1. M. Fitzner, A. Šošić, A. Hopp, et al., unpublished 2024 github.com/emdgroup/baybe.
2. B.J. Shields, J. Stevens, J. Li, et al., Nature 2021, 590, 89–96.

 Joscha Hoche and Marcel Müller