Sasha Rush (ICLR) here is my taxonomy of these methods and some works till 2021 I was tracking

I think this is based on an unfair comparison of 'filtering out invalid SMILES' versus 'retaining all SELFIES'.
Filtering out invalid SMILES improves the model's performance because they have low likelihoods. In fact, one can also filter out SELFIES with low likelihoods.

Penzai’s Colab integration is just 🤯

I guess the “rule of four” is a result of: 1) the distribution of materials among space groups; 2) the distribution of atoms in Wyckoff positions; 3) the multiplicities of Wyckoff positions. 1) and 2) are determined by inter-atomic interactions while 3) is fixed by group theory.

Google 翻译

😅中译英——英译中后的效果

“autoregressive” versus “transportation”would be the right way to contrast

Goodfellow talking about controlling a Tokamak with RL at Mila.

Thrilled to officially be a part of Team W1LD and introduce my new signature off-string yo-yo, 'Boundless'!

Introducing #XtalNet by Guolin Ke and team: E2E crystal structure prediction from PXRD data with contrastive learning and #diffusion based conditional crystal structure generation. paper: arxiv.org/abs/2401.03862 dataset：drive.google.com/drive/folders/… #AI4Sci #GenAI #Crystallography

Study on the electronic properties and superconductivity of bilayer La3Ni2O7 cpl.iphy.ac.cn/10.1088/0256-3…

The tensor-network message-passing method reduces the error in the calculation of local observables by several orders of magnitude compared with state-of-the-art techniques go.aps.org/3IYvItD

😀For those curious about the math of rectified flow, the paper is here: arxiv.org/abs/2209.03003

Unlocking the secrets of high-temperature superconductor La3Ni2O7 by integrating density functional theory and a sophisticated bi-layer Hamiltonian cpl.iphy.ac.cn/10.1088/0256-3…