Research
Minimizing data requirements for chemical and material property prediction via cluster-based training set selection
Nature · By Quinn M. Gallagher ·
The Brief
- A survey of 100 algorithms across 60 chemical and materials property prediction tasks was conducted.
- Cluster-based selection of training data outperformed active learning in the study.
- The paper is published in Nature Computational Science (doi:10.1038/s43588-026-01076-9).
Why it matters
The findings suggest that choosing training sets by clustering can reduce the data needed to train property prediction models, which can lower the cost of generating chemical and materials data.
Who’s affected
- researchers in computational chemistry and materials science
- machine learning practitioners working on chemical and materials property prediction
- #machine learning
- #materials science
- #chemistry
- #active learning
- #training data selection
Nature Computational Science, Published online: 18 November 2026; doi:10.1038/s43588-026-01076-9 A survey of 100 algorithms across 60 chemical and materials property prediction tasks finds that cluster-based selection of training data outperforms active learning, and a…