Cédric Bourgès
(International Center for Young Scientists, National Institute for Materials Science
)
;
Guillaume Lambard
(Center for Basic Research on Materials/Data-driven Materials Research Field/Data-driven Materials Design Group, National Institute for Materials Science
)
;
Naoki Sato
(Research Center for Materials Nanoarchitectonics (MANA)/Nanomaterials Field/Thermal Energy Materials Group, National Institute for Materials Science
)
;
Makoto Tachibana
(Research Center for Materials Nanoarchitectonics (MANA)/Nanomaterials Field/Thermal Energy Materials Group, National Institute for Materials Science
)
;
Satoshi Ishii
(Research Center for Materials Nanoarchitectonics (MANA)/Nanomaterials Field/Optical Nanostructure Team, National Institute for Materials Science
)
;
Takao Mori
(Research Center for Materials Nanoarchitectonics (MANA)/Nanomaterials Field/Thermal Energy Materials Group, National Institute for Materials Science
)
Description:
(abstract)The active learning by machine learning and Bayesian optimization pipeline (ALMLBO) is a new tool rising in the experimental material development. The pipeline is a general framework that comprises: a “learning” step based on past experiments and during which a statistical modelization of a system {process parameters/composition, targeted properties} is attempted by using experimental data to train a machine learning model that capture relationships between parameters and properties; And an “active” step in which a set of experimental actions, derived from the learned model of the system, are performed and are supposed to bring the system closer to an objective. Bayesian optimization leverages this model-building to guide the choice of parameter sets. Multiple learning-acting cycles constitute an active learning pipeline.
In the present case, the use of the ALMLBO has been developed on a strategic material, the kesterite [3], to develop a process control and a fine composition adjustment as both key factor for obtaining a superior thermoelectric (TE) property. On the first hand, it can reduce the number of experiments required to find the ideal set of process parameters or composition tunning that improve TE properties and, on the second hand, propose statistical relationships between the process parameters and the targeted physical properties. Notably, the latter advantage supports the establishment of dependencies which could appear pertinent to the understanding of a physicochemical system like in the present study.
Rights:
Keyword: Kesterite, Machine learning, Process, Thermoelectric, Ceramic
Conference: THE 11th INTERNATIONAL WORKSHOP ON ADVANCED MATERIALS SCIENCE AND NANOTECHNOLOGY (2024-09-22 - 2024-09-25)
Funding:
Manuscript type: Author's original (Submitted manuscript)
MDR DOI: https://doi.org/10.48505/nims.4910
First published URL:
Related item:
Other identifier(s):
Contact agent:
Updated at: 2025-04-10 21:55:17 +0900
Published on MDR: 2024-10-30 16:30:25 +0900
Filename | Size | |||
---|---|---|---|---|
Filename |
Revised Manuscript-ActaMater_20240327.docx
(Thumbnail)
application/vnd.openxmlformats-officedocument.wordprocessingml.document |
Size | 3.85 MB | Detail |