Reproduce “Omics-scale polymer computational database transferable to real-world artificial intelligence applications” on OpenResearch

About this paper

Ryo Yoshida, Yoshihiro Hayashi, Hidemine Furuya, Ryohei Hosoya, Kazuyoshi Kaneko, Hiroki Sugisawa, Yu Kaneko, Aiko Takahashi and 98 more

The RadonPy consortium developed PolyOmics, an "omics-scale" computational database containing over 7.3 million entries for 62 properties across 10^5 distinct polymers, generated through 100 million node-hours on the Fugaku supercomputer. This work demonstrates power-law scaling in Simulation-to-Reality transfer learning, showing that pre-training AI models on this large computational dataset significantly enhances their ability to predict real-world polymer properties.

Key insights

• Omics-scale computational databases can serve as foundational resources for AI-driven materials discovery by enabling effective Sim2Real transfer learning, even with inherent simulation imperfections.
• The generalization performance of AI models in materials science improves predictably with increasing computational dataset size, validating a power-law scaling relationship for ultralarge-scale data.
• Computational datasets can explore vast, previously unsynthesized polymer chemical spaces, leading to the identification of novel material candidates that surpass empirically known property trade-offs.

Summary by alphaXiv.