NASA and IBM's open-source lunar AI model matches or beats Earth-trained baselines
NASA-IBM Lunar Foundation open-Source Geospatial AI Model
NASA and IBM Research have released the NASA-IBM Lunar Foundation Model, an open-source AI pretrained from scratch on SomBench, a multimodal dataset of nearly two million co-registered bundles across 11 modalities. USRA planetary scientist Rachel Slank helped connect lunar science priorities with model development. Across crater detection, irregular mare patch segmentation, and polar ice prospectivity regression, the pretrained model matched or outperformed ImageNet-based and non-pretrained baselines, showing strong label efficiency. Model, fine-tuning code, and benchmarks are on Hugging Face.
One of the biggest challenges was bringing together lunar datasets that span very different instruments and spatial resolutions (1m to 20km per pixel) while still preserving the scientific value of each dataset.
- petcat
Good to see actual open-source models instead of these so-called "open weight" models which are still just inscrutable binary blobs that don't even provide a high-level catalog or really any description whatsoever of what went into the training data or the methods and tools used to train them.
- WalterGR
Dupe of https://news.ycombinator.com/item?id=49697549
“A rough guide for going back to the Moon” (ibm.com) - 162 comments