A Bayesian optimized spectral recommender system with dynamic human-guided targets for physics discovery

The combination of computational design, optimization, and machine learning with large scale microscopic data analysis, has led to the automated experiments towards rapid discovery of physics. In finding optimal parameters for desired material properties where the functional maps between them is unknown or expensive, a Bayesian optimization (BO) is well suited due to handling of any black-box objective functions (numerical functional forms are not required) and adaptive sampling technique for minimal expensive evaluations for convergence. However, the BO applications are generally limited to a prior set target material property, and in experimental analysis, often it is very complex to priorly know such properties. Here, we have extended the BO application where the user sequentially learns and update the target properties as the optimization progress, and thereby build a BO-based spectral recommendation system (BO-SRS). As the BO progresses, the recommender first input a spectral, then evaluates the quality of the spectral through a custom build objective function based on the user (domain expert) decision and set/updated current target material properties, then the BO samples the next best location towards spectral of similar material properties as the current target. We demonstrate the performance of BO-SRS workflow on both piezoresponse force spectroscopy as well as current-voltage (I-V) spectroscopy on complex oxide samples, with multiple users and different user goals. This stated automated human-augmented approach allows for rapid device parameter selection for a complex priorly unknown map between the application and the respective material properties.