Pallete image processing models (e.g. JPEG restoration, inpainting), adapted to act as a surrogate model for density functional theory (DFT) relaxation of crystal structures and structure prediction.
MIT License
0
stars
0
forks
source link
Figured it was better to get this idea public, so people can say they're already doing it/done, or be aware before starting #1
Like with many other state-of-the-art algorithms, there seems to be several groups trying to implement it in materials-/chem-informatics confidentially for up to a year or two prior to publication (or posting a preprint). Not that it's bad to have the similarities, differences, and some competition that arises, but personally I tend to prefer working on unique projects and quickly turning them around. Diffusion models are a big deal, and I wouldn't be surprised if there are already 3-5 major materials-/chem-informatics groups thinking about or working on adapting them, similar to what's happened with transformers, VAEs, and GANs.
If you come across this and you've been thinking of or are already working on an implementation of diffusion models for materials/cheminformatics, would love to chat and figure out how we can divide and conquer / learn from each other to each make unique contributions.
Like with many other state-of-the-art algorithms, there seems to be several groups trying to implement it in materials-/chem-informatics confidentially for up to a year or two prior to publication (or posting a preprint). Not that it's bad to have the similarities, differences, and some competition that arises, but personally I tend to prefer working on unique projects and quickly turning them around. Diffusion models are a big deal, and I wouldn't be surprised if there are already 3-5 major materials-/chem-informatics groups thinking about or working on adapting them, similar to what's happened with transformers, VAEs, and GANs.
If you come across this and you've been thinking of or are already working on an implementation of diffusion models for materials/cheminformatics, would love to chat and figure out how we can divide and conquer / learn from each other to each make unique contributions.