Sharing new breakthroughs and artifacts supporting molecular property prediction, language processing, and neuroscience

Abstraction: Meta FAIR releases OMol25 dataset, UMA model, Adjoint Sampling, and brain language study

Key points:

• OMol25 (Open Molecules 2025): largest DFT dataset for molecular chemistry, built with ORCA 6.0.1, required 6 billion core hours; configurations up to 10x larger than prior datasets (vs. 20-30 atoms previously)
• UMA (Universal Model for Atoms): ML interatomic potential trained on 30+ billion atoms across all Meta open science datasets (Open Catalyst, Open DAC, Open Materials, OMol25)
• Adjoint Sampling: reward-based generative modeling algorithm that trains without any data, using iterative self-refinement guided by a reward model; useful for computational chemistry and fine-tuning
• Neuroscience study with Rothschild Foundation Hospital: analyzed neural recordings from 7,000+ electrodes in 40+ epilepsy patients to map language representation development from age 2 onward
• Brain-AI parallel: language representations in developing brains spontaneously mirror those of LLMs like wav2vec 2.0; children acquire language with ~1,000x fewer words than LLMs

Connections: Meta · Meta Fair · Molecular Property Prediction · Large Language Models · Generative Models

Source: https://ai.meta.com/blog/meta-fair-science-new-open-source-releases/?utm_source=twitter&utm_medium=organic+social&utm_content=video&utm_campaign=fair