Download file (4.15 MB)
Download as zip (3.73 MB)

Collection

Citation

Description:

In this study, we demonstrate that the scaling law of simulation-to-real (Sim2Real) transfer learning holds for several machine learning tasks in materials science. Through three different prediction tasks for polymers and inorganic material systems, it was observed that the prediction error on real systems decreased monotonically with increasing the size of computational data according to a power law. Observing the scaling behavior offers various insights for advancing database development, such as determining the sample size necessary to achieve a desired predictive performance and a quantitative measure of the database’s potential value in real-world applications. Additionally, it aids in identifying equivalent sample sizes for physical and computational experiments and guiding the design of data production protocols principled the Sim2Realtransferability and scalability to downstream real-world tasks.

Rights:

• Creative Commons BY Attribution 4.0 International
  

Keyword: molecular dynamics simulation, scaling law of simulation- to-real, Sim2Real, transfer learning, PoLyInfo

Date published: 2025-05-24

Publisher: Springer Science and Business Media LLC

Journal:

• npj Computational Materials (ISSN: 20573960) vol. 11 issue. 146

Funding:

• Ministry of Education, Culture, Sports, Science and Technology hp210264
• MEXT | JST | Core Research for Evolutional Science and Technology JPMJCR19I3, JPMJCR22O3, JPMJCR2332
• MEXT | Japan Society for the Promotion of Science 19H05820, 19H01132
• MEXT | Japan Society for the Promotion of Science 23K19980
• MEXT | Japan Society for the Promotion of Science 22K11949

Manuscript type: Publisher's version (Version of record)

MDR DOI:

First published URL: https://doi.org/10.1038/s41524-025-01606-5

Related item:

Other identifier(s):

Contact agent:

Updated at: 2025-06-13 16:30:25 +0900

Published on MDR: 2025-06-13 16:20:59 +0900