LunarFM: a multimodal representation of the Moon's surface

Recent decades have seen a revival of lunar exploration, driven by the dual goals of establishing a sustained human presence and assessing the Moon’s resource potential; however, current resource mapping remains fragmented due to heterogeneous, multi-instrument datasets that expose fundamental limitations in existing analysis pipelines. We introduce LunarFM, the first foundation model designed to unify heterogeneous lunar remote-sensing observations from multiple instruments in a shared representation space, integrating six data modalities from three missions and comprising 18 input channels. It is trained via multimodal self-supervised pretraining informed by the data's spatial and physical structure. LunarFM is a fully open-source framework, consisting of: (1) a machine-learning-ready dataset of co-registered multimodal chips spanning 0.5°$\times$0.5° of lunar latitude and longitude; (2) a multimodal foundation model based on a multi-masked autoencoder; (3) a companion embedding dataset providing a joint 768-dimensional representation of lunar surface properties; and (4) a suite of downstream applications, including similarity search and few-shot resource (e.g., titanium dioxide) mapping. By enabling accessible, reproducible, and scalable multimodal analysis, LunarFM supports improved lunar scientific investigation and more robust resource prospecting. We discuss design choices, limitations, and evaluation strategies relevant to deploying foundation models in planetary science workflows, thereby facilitating data-driven site characterisation and resource assessment for future exploration missions. LunarFM and all corresponding datasets will be made publicly available in the near future.