Probing out-of-distribution generalization in machine learning for materials

Kangming Li, Andre Niyongabo Rubungo, Xiangyun Lei, Daniel Persaud, Kamal Choudhary, Brian DeCost, Adji Bousso Dieng, Jason Hattrick-Simpers

Keywords: machine learning, materials science, out of distribution, graph neural network

TL;DR: We evaluate 700+ out-of-distribution (OOD) tasks in materials science and show that current OOD benchmarks overestimate AI capabilities: they test interpolation, not true extrapolation, masking the limits of scaling and generalizability.

Confirmation Of Submission Requirements: I submit a previously published paper. It was published in an archival peer–reviewed venue on or after September 8th 2024, I specify the DOI in the field below, and I submit the camera-ready version of the paper.

DOI: https://doi.org/10.1038/s43246-024-00731-w

Submission Number: 25