BOPE-GPT, Preference Exploration with the curious AI chemist

Abstract

For the Bayesian Optimization Hackathon for Chemistry and Materials (27-28 March 2024), our team finished and presented the project on BOPE-GPT, Preference Exploration with the curious AI chemist.

Bayesian optimization with preference exploration (BOPE) is designed for Bayesian optimization of expensive-to-evaluate experiments, where the response surface function of the experiment generates vector-valued outcomes over which a decision-maker (DM) has preferences. In this project, we combined the BOPE pipeline with a language model, using the language model as the decision-maker. We worked on a multi-pbjective optimization problem (optimal production of the Fischer-Tropsch reaction) with this pipeline. Furthermore, we built an app for easy LLM-based BOPE.

I mainly worked on the data preprocessing and Bayesian optimization pipeline (Single-objective BO, Multi-objective BO and BOPE) building.

See project description and our Git-hub page for more details.