Diffusion Autoencoders with Perceivers for Long, Irregular and Multimodal Astronomical Sequences

We introduce a perceiver-based diffusion autoencoder (daep) for learning bottleneck representations from irregularly sampled long sequences across multiple modalities. Our approach encodes each modality with a perceiver encoder, optionally applies late fusion for multimodal data, and decodes with modality-specific diffusion decoders. Unlike standard $\beta$-VAEs, daep can produce more regular latent spaces and capture finer details in reconstructions, as demonstrated on high-resolution variable star spectra, ZTF supernova light curves and spectra, and Galaxy10 images. We further show that the model supports cross-modality generation and handles missing modalities through modality dropping. These results highlight daep’s potential for downstream tasks such as classification, clustering, and generative modeling of complex scientific data.