Bayesian Optimization for Polymer Blend Phase Boundary Detection: Benchmarking with a Simulated Ground Truth Model

Autonomous experiments, particularly Gaussian Process-informed Bayesian Optimizations, show great potential for accelerating scientific research and improving resource efficiency. However, the complexity of various parameters and decisions within these algorithms necessitates thorough testing before application to experimental systems. In this study, we utilized data from a previous paper by our group (Newby et al., 2000) to reconstruct the lower critical solution temperature (LCST) phase diagram of a poly(methyl methacrylate) and poly(styrene-ran-acrylonitrile) blend. This reconstruction enabled us to simulate different acquisition functions that balance exploration–sampling untested regions of phase space–and exploitation–focusing on areas of interest, such as the phase boundary–to recommend optimal composition and temperature combinations for sample preparation and characterization. We also developed decision policies to enhance efficiency beyond single-point serial preparation, including the strategy of preparing multiple compositions at the same temperature. Through iterative simulations, we identified frameworks for autonomous experimentation that are generalizable to unknown systems, paving the way for deployment in laboratory experiments.