You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I think it would be a good idea to lessen the focus from CDVAE and to also include a few other notable generative models. While CDVAE was important as one of the first demonstrations of generative models creating periodic lattices, it also has some drawbacks including difficulty in installation, mode collapse with respect to symmetry, and the developers themselves have moved to other models including conditionally generative models.
I would recommend a small change to include the following
Generative models have become important in materials informatics as a way to expand the pallette of potential candidate materials beyond those already known in crystal structure databases. Truly new materials will allow researchers to go beyond screening to discovery. There are a number of important models that have emerged. For example, our own group has released xtal2png and CrysTens representations and tested them with diffusion models and GANs of various flavors. There is also FTCP from MIT which was deployed with a VAE model. More recently there is MatterGen which is an extension of CDVAE but is now property-conditioned. There have also been numerous reinforcement learning approaches such as ORGAN some of which are image-to-image inspired in order to generate compounds or molecules as variants of known molecules.
The text was updated successfully, but these errors were encountered:
I think it would be a good idea to lessen the focus from CDVAE and to also include a few other notable generative models. While CDVAE was important as one of the first demonstrations of generative models creating periodic lattices, it also has some drawbacks including difficulty in installation, mode collapse with respect to symmetry, and the developers themselves have moved to other models including conditionally generative models.
I would recommend a small change to include the following
Generative models have become important in materials informatics as a way to expand the pallette of potential candidate materials beyond those already known in crystal structure databases. Truly new materials will allow researchers to go beyond screening to discovery. There are a number of important models that have emerged. For example, our own group has released xtal2png and CrysTens representations and tested them with diffusion models and GANs of various flavors. There is also FTCP from MIT which was deployed with a VAE model. More recently there is MatterGen which is an extension of CDVAE but is now property-conditioned. There have also been numerous reinforcement learning approaches such as ORGAN some of which are image-to-image inspired in order to generate compounds or molecules as variants of known molecules.
The text was updated successfully, but these errors were encountered: