emdgroup/baybe

What it solves

BayBE (Bayesian Back End) is designed to find optimal parameter configurations within complex search spaces. It is particularly useful for real-world optimization problems where experiments are costly or time-consuming, such as chemical reaction optimization, materials formulation, 3D shape optimization, and hyperparameter tuning for virtual simulations.

How it works

BayBE uses Bayesian Design of Experiments to balance the exploration of unknown regions and the exploitation of known high-performing regions of a parameter space. The workflow follows a recommendation loop: the user defines a search space and an objective, and BayBE recommends the next set of configurations to test. Once the user provides the results (measurements), BayBE updates its model to refine future recommendations.

Who it’s for

It is intended for researchers and engineers in industry and academia who need to optimize physical or virtual experiments, particularly those working with chemical substances or complex hybrid search spaces.

Highlights

• Hybrid Search Spaces: Supports both continuous and discrete parameters in a single space.
• Chemical Encodings: Includes built-in support for chemical fingerprints (e.g., MORDRED) via SubstanceParameter to improve optimization efficiency over simple one-hot encoding.
• Flexible Optimization: Offers various strategies including active learning, bandit models for AB testing, and Pareto optimization for multiple conflicting targets.
• Advanced Workflows: Supports asynchronous campaigns, transfer learning from previous campaigns, and backtesting for performance evaluation.