IBM and NASA have developed an open source copy of artificial intelligence designed more exactly to allow scientists to scrutinise the Moon and used the observations that have been on the Moon for decades as a tool that can now be re-used to explore the lunar surface for resources.
The NASA-IBM Lunar Foundation Model was released to the public on September 10 as one of the first open source foundation models developed in particular for lunar science. The model integrates various kinds of lunar observations, providing an opportunity for patterns to emerge that would be otherwise difficult to identify when looking at data sets individually.
This project is one response to an increasing problem in lunar research. NASA and other space agencies have amassed 30 years of data from spacecraft and scientific experiments into vast collections. Making good use of those data has traditionally meant painstakingly interpreting those archives from scratch or designing mission-specific machine-learning systems (.
The new model is more di fferent, in that it integrates two or more observations from multiple instruments at contrasting resolutions. Researchers at IBM and NASA assembled a common, machine-learning-ready, multi-resolution lunar dataset of over 30 spatially registered data layers captured by nine instruments and from four different lunar missions. The dataset comprises values provided by GRAIL and the LRO mission beyond the data provided by the SELENE/Kaguya mission, a Japanese mission.
A notable application of the model is in detecting where deposits of water ice may exist. Permanently shadowed regions near the lunar poles are hard to observe, as sunlight cannot reach them directly. Such regions are considered likely to host deposits of frozen water that may later prove to be a useful resource.
As IBM and NASA, the model is able to fuse sources of information like topography and temperature to assist in finding promising regions for further study. During ice detection tests, IBM has further noted, the model decreased the margin of error against the state-of-the-art SwinV2 transformer that was trained for the same task by 22%.
The model may also be useful for assisting in identifying lunar craters. (e.g. crater maps can be used to study the geological history of the Moon and high-quality terrain data could be used to aid the selection of future landing sites and the planning of landing sites and infrastructure.) Based on IBM and NASA, the model delivered equivalent performance as state-of-the-art systems in meter-scale crater mapping and improved performance at larger contexts scales.
Another area of investigation is the volcanic history of the Moon. With this model, it would be possible to analyze the lunar geology, like formations and spots of old lava, for further insight into the development of the lunar surface over billions of years.
Open source is one of the biggest advantages of the project. NASA mentions that the model is hosted by Hugging Face publically, and the full code repository resides on GitHub. This allows scientists, academia and any other developers to evaluate the model, modify it for their own purposes and stand-alone applications.
This lunar model further develops the partnership established between IBM and NASA. Their work together had earlier resulted in the Prithvi family of open-source geospatial models, designed for observation of the Earth, and the Surya heliophysics and space-weather model. This new lunar system extends that work ‘upward’ to a third planetary setting: