Bayesian optimization, an active learning method for optimising experimental parameters

LAMBARD Guillaume; BAJAN Christophe Marie Olivier

口頭発表 Bayesian optimization, an active learning method for optimising experimental parameters

LAMBARD Guillaume (Center for Basic Research on Materials/Data-driven Materials Research Field/Data-driven Materials Design Group, National Institute for Materials Science) ; BAJAN Christophe Marie Olivier (Center for Basic Research on Materials/Data-driven Materials Research Field/Data-driven Materials Design Group, National Institute for Materials Science)

Download file (4.01MB)
Download as zip (3.63MB)

コレクション

引用

LAMBARD Guillaume, BAJAN Christophe Marie Olivier. Bayesian optimization, an active learning method for optimising experimental parameters. https://doi.org/10.48505/nims.5198

(BibTeX)

説明:

(abstract)

A supervised machine learning model (SML) is a method of artificial intelligence (AI) that enables
the estimation of the value of various properties (called ”target” values) from the values of other
properties (called ”features”). SML is a particularly valuable tool in situations where it is not possible to calculate the target values using an analytical function, and more over in cases where the target values are challenging or time-consuming to determine. Such an approach can be applied to values obtained through numerical calculation or experimental measurement. In general, SML requires a big amount of data, typically ranging from hundreds to several thousand data points. Applying this method to experimental synthesis can be difficult, as the synthesis of more than a hundred compounds may prove impossible. Bayesian optimization (BO) represents a specific type of SML within a narrower category of models known as active learning. The objective is to identify an optimal value for the target properties with a minimum amount of data. To achieve this, we start with a modest amount of data and incrementally increase it, selecting the most promising potential values of features through a combination of Gaussian process and SML. The key strength of BO lies in its acquisition function, which guides the selection of the next point to evaluate. This function intelligently trades off between exploring uncertain regions and exploiting promising areas, allowing BO to converge to the optimum with less data than to other optimization methods. In this seminar, I will present the fundamental concepts of BO and demonstrate its practical application using MADGUI, a fully graphical interface developed by C. Bajan and G. Lambard at NIMS. This tool makes BO accessible to researchers without extensive programming experience.

権利情報: