experimental-design/bofire

What it solves

BoFire is a comprehensive framework for experimental design and Bayesian optimization (BO), specifically tailored for real-world physical experiments. It addresses the challenges of optimizing complex systems where experiments are costly or time-consuming, such as in chemical reactions, pharmaceutical formulations, and digital twins.

How it works

The framework uses an iterative "ask-and-tell" loop to find optimal parameters. Users define a search space (Domain) consisting of continuous, discrete, or categorical inputs and desired outputs. The system then suggests new candidates for testing based on a surrogate model (built on BoTorch) and an acquisition function. After the user performs the experiment and provides the results back to the system, the model is updated to refine future suggestions.

Who it’s for

It is designed for researchers, data scientists, and engineers in the chemical and pharmaceutical industries, as well as academic partners who need to optimize physical experiments.

Highlights

• Mixed Parameter Spaces: Supports continuous, discrete, and categorical inputs and outputs.
• Chemical Intelligence: Includes built-in chemical encodings and kernels to improve performance when optimizing molecules.
• Constraint Handling: Supports specific, generic, and black-box output constraints.
• LLM Integration: Features an LLMStrategy that uses large language models to propose candidates, which is particularly useful for "cold starts" where prior written knowledge exists.
• REST API Ready: Built-in serialization capabilities make it easy to integrate into web services.