# Fileset

[Poster_GDR_IAMAT_NIMS_2025.pdf](https://mdr.nims.go.jp/filesets/22e2a357-5101-4959-8691-977e9f876e14/download)

## Creator

[LAMBARD Guillaume](https://orcid.org/0000-0003-0275-4079), [BAJAN Christophe Marie Olivier](https://orcid.org/0009-0008-1433-9618)

## Rights

Copyright 2026 LAMBARD Guillaume, BAJAN Christophe Marie Olivier

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

## Other metadata

[Active Learning for Materials Science application.](https://mdr.nims.go.jp/datasets/ed3419da-8bd5-402b-bca4-cbcd963c6aa1)

## Fulltext

Active learning to accelerate research in material scienceC. Bajan1,* & G. Lambard1,* 1. Center for Basic Research on Materials (CBRM) Data-driven Materials Design Group, National Institute for Materials Science (NIMS), Tsukuba, Japan* : BAJAN.Christophe@nims.go.jp ; LAMBARD.Guillaume@nims.go.jpAccelerating materials discovery is a critical challenge in both academic and industrial research. Traditional trial-and-error approaches are often slow and resource-intensive, especially when exploring vast compositional and processing spaces. Active learning revolutionizes materials discovery by strategically guiding experimentalsynthesis through machine learning algorithms. In this work, we demonstrate how we used machine learning, powered by active learning algorithms, can strategically guideexperimental synthesis to maximize information gain and efficiency. In this study, we applied a Bayesian optimization framework to select the most informative nextexperiments in the development of two targets properties. This approach enabled us to significantly reduce the number of experiments required to achieve targetproperties, compared to conventional grid search methods. To facilitate the adoption of these techniques by researchers and industry partners, we developed MADGUI, anintuitive, open-source graphical user interface built with Streamlit. MADGUI streamlines the active learning workflow, allowing users to visualize predictions, monitoruncertainty, and interactively select new candidates for synthesis. Our collaborative results showcase the practical impact of integrating active learning into real-worldmaterials research, highlighting substantial gains in efficiency and discovery speed. This work underscores the transformative potential of machine learning-drivenexperimentation for accelerating innovation in materials science.Active learningIn this study, active learning was implemented across 11 experimental parameters, creating a vastsearch space of approximately 10¹⁸ possible combinations, far beyond what could be exhaustivelyexplored. The workflow operates in iterative cycles: after each round of experiments, theexperimentalists provide newly measured data along with updated parameter boundaries. This data isthen pre-processed to meet the requirements of the machine learning algorithm. Some linearcorrelation and prediction model are applied to analyze the data and Bayesian Optimization proposethe most informative next experiments. These suggested parameter sets are then returned to theexperimentalists for synthesis and measurement, and the cycle repeats, enabling efficient explorationand accelerated materials discovery.- G’(0.1%)-G’(50%)- Electrical PowerFeatures importanceA well-performing prediction model enables us to quantify the influence of each parameter onthe optimization of target properties. By analyzing feature importance, we can identifyparameters that have negligible impact on the target values, allowing us to reduce thedimensionality of the problem and streamline the experimental search space. Conversely,parameters with significant influence are prioritized, enabling us to focus subsequent studies onthe variables that most strongly affect material performance. This targeted approach enhancesboth the efficiency and effectiveness of the optimization process.Prediction modelAfter several iterations, the active learning process yields a comprehensiveunderstanding of the parameter space, allowing us to efficiently map regions associatedwith optimal material properties. By leveraging the algorithm’s balance betweenexploration (sampling uncertain regions) and exploitation (focusing on promisingcandidates), we progressively refine our predictive models. This approach enablesaccurate identification of parameter sets that meet or exceed target performancecriteria, greatly enhancing our ability to predict and achieve desired material outcomeswith fewer experiments.Prediction model for the two targets of this studyData AnalysisFor the latest iterations, we reduced the number of parameters from 11 to 5, focusingon those most relevant to our target properties. Using a prediction model trained onthe experimental dataset, we predicted target values based on these selectedparameters. This analysis revealed the trade-offs between the two key target propertiesand allowed us to identify the two most influential parameters for each target. Notably,we observed a clear linear correlation between rpm and electrical power, while mixingtime exhibited a negative linear correlation with the G’ factor, providing valuableinsights into the underlying relationships within the system.Overall, in this study , the machine learning focus on the diminution of both targets at thesame time resulted in an improvement of 46.3% with only 10 iterations, 11 parameters and abetter understanding of the two main properties impact on the targets properties. Slide 1