SinAE: Single Architecture Autoencoder for Cross-Domain Atomic Systems

Traditional generative models for atomic systems like molecules, crystals, and proteins are typically domain-specific, leading to fragmented pipelines and data scarcity challenges. A new model called SinAE addresses this by employing a single, unified Transformer-based flow-matching autoencoder. This architecture avoids specialized graph or equivariant operators, instead shifting the burden of fine-grained geometry reconstruction to an iterative flow-matching decoder.This design allows SinAE to achieve highly accurate, near-lossless reconstruction across diverse atomic domains. The model's shared atomic latent space supports a standard Diffusion Transformer prior, demonstrating strong performance on various generation benchmarks. Crucially, joint training across molecules and crystals shows direct evidence of beneficial cross-domain transfer, highlighting the efficiency gains from a unified approach.