MI-6 Ltd., in collaboration with a research group from National Yang Ming Chiao Tung University, has conducted a study on enhancing data quality management and analytical reliability within the field of Materials Informatics (MI). The results of this research have been published in the international academic journal Oxford Open Materials Science.
While the advancement of MI has accelerated materials development through machine learning, significant challenges remain. Structural inconsistencies and variations in data quality often compromise model reproducibility and prediction stability. In particular, raw experimental data from laboratory settings often possesses diverse structures—where compositional information and processing conditions are recorded in non-standardized formats—which serves as a primary cause for reduced reliability in predictive modeling.
In this study, the group developed an encoder-decoder mechanism to transform heterogeneous experimental data into representations suitable for machine learning. Additionally, they proposed a unique quality metric, the "R:M ratio," which combines explanatory power with error magnitude.
Evaluation using experimental datasets confirmed that filtering based on the R:M ratio contributes to the stability of predictions. Conversely, the results also suggested that the effectiveness of this metric depends on the characteristics of the dataset, indicating the need for further validation that incorporates physical constraints.
This achievement presents a criteria for objectively evaluating the reliability of predictive models by integrating the MI workflow from data management to optimization. It is expected that these results will promote the standardization of processes—from experimental data accumulation to model construction—and contribute to more robust and certain discoveries of novel materials.
Paper Information
- Title: The exploration of enhanced methods in data science and applications in materials science (リンク)
- Journal: Oxford Open Materials Science
- DOI:https://doi.org/10.1093/oxfmat/itaf022
- Authors ※Affiliations are based on the information at the time of publication.
- Raymond Wu, Susumu Otsuki (MI-6 Ltd.)
- Sirirat Sae Lim, Haishang Wu (National Yang Ming Chiao Tung University)
