New framework AtomBench standardizes AI model evaluation for crystal reconstruction

Researchers have developed AtomBench, a new framework for evaluating generative crystal reconstruction models, particularly for conventional superconductors. The framework allows for standardized comparisons by ensuring models receive equal crystallographic information during reconstruction. In tests using the JARVIS Supercon-3D and Alexandria DS-A/B datasets, MatterGen demonstrated the best atomic-coordinate reconstruction, while CDVAE excelled at lattice accuracy. The study also found that conditioning on critical temperature did not consistently improve reconstruction fidelity. AtomBench is released as an open-source Python package to encourage community use and benchmarking. AI

IMPACT Standardizes AI model evaluation for materials science, enabling more reliable comparisons of generative models.

RANK_REASON The cluster describes a new benchmarking framework and its application in a research paper. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Alexandria DS-A/B
• AtomBench
• AtomGPT
• continuous corrected RMSD
• FlowMM
• JARVIS Supercon-3D
• Kamal Choudhary
• Kullback-Leibler divergence
• MatterGen
• mean absolute error
• root-mean-squared displacement

• paper
• model release
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →