Multi-Modal Foundation · polymathic-ai
Aion Base
An omnimodal foundation model predicting across 39 astronomical survey data types using masked modelling.
What Aion Base is
Aion Base (AION-1) is a 300-million-parameter scientific foundation model developed by Polymathic AI for astronomy and physics research. It processes and connects heterogeneous observations across major astronomical surveys, unifying 39 distinct data types that include multi-band imaging, optical spectra, astrometry, and photometric catalogues.
Astronomical research routinely requires synthesising disparate measurements from multiple ground and space observatories. Aion Base addresses this fragmentation by framing observation analysis as an any-to-any multimodal problem, enabling researchers to project multi-survey observations into shared representations or infer missing observational modalities from available inputs.
How it works
The model uses an encoder-decoder Transformer architecture consisting of 12 encoder blocks and 12 decoder blocks, with a hidden dimension of 768 and 12 attention heads. It is trained via multimodal masked modelling (4M) across major astronomical survey datasets, including the Legacy Survey, Hyper Suprime-Cam (HSC) Wide, Gaia, the Sloan Digital Sky Survey (SDSS), and the Dark Energy Spectroscopic Instrument (DESI).
Astronomical inputs—such as image pixel tensors, spectral fluxes, or tabular catalogue entries—are transformed into discrete tokens using survey-specific codecs. The encoder maps these tokens into contextual representations, while the decoder can generate target modalities (such as spectroscopic redshift or alternative survey bands) specified by modality target masks.
Key facts
- Architecture
- Encoder-Decoder Transformer (12 encoder blocks, 12 decoder blocks)
- Parameter count
- 300M
- Hidden dimension
- 768
- Attention heads
- 12
- Training objective
- Multimodal Masked Modelling (4M)
- Supported modalities
- 39 astronomical data types across imaging, spectra, catalogues, and photometry
- Pre-training datasets
- Multimodal Universe (Legacy Survey, HSC, Gaia, SDSS, DESI)
What you can use it for
- Redshift Prediction
- Estimate spectroscopic redshift directly from multi-band survey imaging or photometric measurements by setting redshift as the target generation modality.
- Cross-Survey Representation Learning
- Extract unified latent embeddings from combinations of images, spectra, and catalogue parameters for downstream astronomical classification, clustering, or anomaly detection.
- Cross-Modality Generation
- Predict unobserved astronomical measurements, such as missing photometric fluxes, extinction values, or morphological shape components, conditioned on available survey data.
Run it here
We list this model but cannot run it yet: unsupported task: any-to-any.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Unifies 39 diverse astronomical data types spanning imagery, spectroscopy, astrometry, and tabular catalogues within a single model.
- Supports flexible any-to-any inference, enabling both cross-modal prediction and embedding extraction.
- Pre-trained on comprehensive observational data from five major astronomical surveys (Legacy Survey, HSC, Gaia, SDSS, and DESI).
Watch out for
- Restricted to the 39 pre-defined astronomical data types and token schemas from supported surveys.
- Inputs must be converted into survey-specific token formats using companion codecs before ingestion by the transformer.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · mit
The mit licence carries terms worth reading before you depend on it.
Questions
Can I use Aion Base commercially?
Check the licence before commercial use. The mit licence carries terms worth reading before you depend on it.
Which surveys are supported by Aion Base?
The model natively supports data products from the Legacy Survey, Hyper Suprime-Cam (HSC) Wide, Gaia, the Sloan Digital Sky Survey (SDSS), and the Dark Energy Spectroscopic Instrument (DESI).
How are different observational formats fed into the model?
Observations are processed through dedicated codecs that convert image tensors, spectral flux arrays, and scalar catalogue values into unified token sequences.
Can the model output both embeddings and specific physical values?
Yes. Researchers can extract 768-dimensional token embeddings for downstream tasks or prompt the decoder with a target mask to predict specific quantities such as spectroscopic redshift.
How to cite it
AION-1: Omnimodal Foundation Model for Astronomical Sciences, arXiv:2510.17960
Written from Aion Base's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.