Smoothing Dark Areas in Molecular Latent Diffusion

Abstract

Latent diffusion for 3D molecular generation depends on a smooth and navigable latent space, yet reconstruction-trained molecular VAEs can contain dark areas that decode to disconnected or chemically invalid structures. This work introduces TopVAE, a topology-optimized VAE that internalizes structural and chemical validity during training, improving robustness and molecular quality without test-time correction.

In brief

A topology-optimized VAE that creates smoother, more chemically valid latent spaces for 3D molecular diffusion.