Not as simple as we thought: a rigorous examination of data aggregation in materials informatics

09 August 2023, Version 1
This content is a preprint and has not undergone peer review at the time of posting.


Recent Machine Learning (ML) developments have opened new perspectives on accelerating the discovery of new materials. However, in the field of materials informatics, the performance of ML estimators is heavily limited by the nature of the available training datasets, which are often severely restricted and unbalanced. Among practitioners, it is usually taken for granted that more data corresponds to better performance. Here, we investigate whether different ML models for property predictions benefit from the aggregation of large databases into smaller repositories. To do this, we probe three different aggregation strategies prioritizing training size, element diversity, and composition diversity. For classic ML models, our results consistently show a reduction in performance under all the considered strategies. Deep Learning models show more robustness, but most changes are not significant. Furthermore, to assess whether this is a consequence of a distribution mismatch between datasets, we simulate the data acquisition process of a single dataset and compare a random selection with prioritizing chemical diversity. We observe that prioritizing composition diversity generally leads to a slower convergence toward better accuracy. Overall, our results suggest caution when merging different data sources and discourage a biased acquisition of novel chemistries when building a training dataset.


Materials Informatics
Machine Learning
Data Aggregation


Comments are not moderated before they are posted, but they can be removed by the site moderators if they are found to be in contravention of our Commenting Policy [opens in a new tab] - please read this policy before you post. Comments should be used for scholarly discussion of the content in question. You can find more information about how to use the commenting feature here [opens in a new tab] .
This site is protected by reCAPTCHA and the Google Privacy Policy [opens in a new tab] and Terms of Service [opens in a new tab] apply.