An Experiment-Aware Bayesian Optimization Workflow for Noisy Mixed-Input Settings

Optimizing expensive black-box objectives over mixed and discretized search spaces is a common challenge in experimental materials science. Bayesian optimization (BO) offers sample-efficient decision-making, but in highly discretized and noisy experimental settings, standard GP-based BO workflows frequently stagnate due to repeated sampling of identical or near-identical conditions. Although alternative surrogate models for mixed-variable BO exist, GP-based surrogates remain most effective for BO in the low-data regime of experimental settings. We therefore present an Experiment-Aware Bayesian Optimization (EABO) workflow that directly targets resampling-induced stagnation while retaining the inherent benefits of GP uncertainty modelling. Rather than replacing GP surrogates, our approach focuses on mitigating this dominant practical failure mode encountered in experimental deployments. The workflow introduces explicit stagnation detection and a dynamic exploration fallback that enforces global exploration when redundant acquisitions occur. We demonstrate its effectiveness on noisy, highly discretized benchmark problems representative of mixed search spaces encountered in experimental materials science. Our work thereby provides a ready-to-use tool for mixed-variable experimental materials design and autonomous laboratories.