USRA Contributes Planetary Science Experience to NASA-IBM Lunar Foundation Model
Open-source synthetic intelligence mannequin combines various lunar datasets to help scientific evaluation of the Moon
The NASA-IBM mannequin reproduces patterns of lunar ice prospectivity (scaled from blue to yellow), proven at 4 places (left) close to the Moon’s pole. Top row: reference ice prospectivity map of Mons Mouton close to the lunar south pole; center row: predictions from the ConvNeXt mannequin; backside row: predictions from the NASA-IBM mannequin. The NASA-IBM mannequin preserves many fine-scale prospectivity patterns within the reference knowledge. Image credit score: NASA/IBM Research
WASHINGTON, D.C., — September 18, 2026. Universities Space Research Association (USRA) contributed planetary science experience, lunar dataset improvement, and scientific analysis to the newly launched NASA-IBM Lunar Foundation Model, an open-source synthetic intelligence (AI) mannequin designed to assist researchers analyze the big, various datasets collected by lunar missions.
Developed by way of a collaboration led by NASA and IBM Research, the NASA-IBM Lunar Foundation Model was pretrained from scratch utilizing SomBench, a multimodal lunar dataset containing practically two million co-registered knowledge bundles spanning 11 modalities and two spatial scales. The mannequin brings collectively complementary details about the lunar floor, together with imagery, topography, illumination geometry, thermophysical properties, mineralogy, radar, gravity, and different geologic and environmental knowledge.
USRA’s contribution to the challenge was supplied by Dr. Rachel Slank, an affiliate scientist with USRA’s Science and Technology Institute, on project at NASA’s Marshall Space Flight Center. She served as a planetary science subject-matter skilled on the NASA-IBM Lunar Foundation Model group. Slank labored throughout each the science and modeling groups, serving to join lunar science priorities and the bodily traits of planetary datasets with selections about mannequin improvement, purposes, and analysis.
The NASA-IBM Lunar Foundation Model was evaluated throughout three downstream benchmarks: crater detection at each regional and meter scales, segmentation of irregular mare patches (IMPs), and regression of lunar polar ice prospectivity. Together, these purposes assess the mannequin’s efficiency throughout a various vary of lunar science challenges, from figuring out influence options and mapping uncommon volcanic landforms to integrating environmental datasets related to the steadiness and potential distribution of polar volatiles.
Across all three benchmarks, the pretrained NASA-IBM Lunar Foundation Model matched or outperformed comparability fashions primarily based on ImageWeb pretraining, in addition to an architecturally an identical mannequin initialized with out lunar pretraining. The research additionally demonstrated significantly sturdy label effectivity in crater detection, suggesting that the representations discovered by way of lunar pretraining can scale back the quantity of task-specific labeled knowledge required for sure purposes.
The multimodal design of the NASA-IBM Lunar Foundation Model permits it to study relationships amongst several types of lunar observations quite than treating every dataset independently. The mannequin was designed to function throughout each regional-scale Wide Angle Camera (WAC) observations and meter-scale Narrow Angle Camera (NAC) knowledge whereas incorporating data comparable to terrain, illumination geometry, and different lunar floor properties.
By releasing the pretrained mannequin, fine-tuning code, and benchmark datasets brazenly, the NASA-IBM Lunar Foundation Model group goals to offer the planetary science and AI communities with a reusable basis for growing new lunar analysis purposes.
A serious element of this work was the collaborative improvement of SomBench, the dataset used each to pretrain the NASA-IBM Lunar Foundation Model and to help standardized analysis of lunar machine studying (ML) purposes.
SomBench accommodates two complementary elements: a core multi-instrument dataset that serves because the pretraining corpus for the NASA-IBM Lunar Foundation Model and a set of utility benchmarks used to guage ML strategies on consultant lunar science issues. The benchmark suite addresses three broad science themes: influence processes, volcanic historical past, and polar volatiles.
As a part of that effort, Slank led improvement of the high-resolution Lunar Reconnaissance Orbiter Camera (LROC) NAC crater benchmark, manually figuring out greater than 49,000 lunar craters. The ensuing dataset makes use of high-resolution lunar imagery along with co-registered digital terrain fashions to guage crater detection at meter-scale decision. She additionally contributed to the broader SomBench datasets and science purposes and supplied in depth scientific evaluation of each the SomBench research and the NASA-IBM Lunar Foundation Model research.
“One of the largest challenges was bringing collectively lunar datasets that span very completely different devices and spatial resolutions (1m to 20km per pixel) whereas nonetheless preserving the scientific worth of every dataset,” stated Dr. Slank. “Because I labored throughout each the science and modeling groups, I may assist make certain these variations had been thought of as the information was introduced collectively and used to coach and consider the muse mannequin. I can’t wait to see what new science researchers will be capable to do with the assistance of the NASA-IBM Lunar Foundation Model!”
The NASA-IBM Lunar Foundation Model and related datasets can be found by way of Hugging Face.
NASA announcement:
https://science.nasa.gov/science-research/artificial-intelligence-lunar-foundation-model/
IBM announcement:
About USRA
Founded in 1969, beneath the auspices of the National Academy of Sciences on the request of the U.S. Government, the Universities Space Research Association (USRA) is a nonprofit company chartered to advance space-related science, know-how, and engineering. USRA operates scientific institutes and services and conducts different main analysis and academic packages beneath federal funding. It engages the college neighborhood and employs in-house scientific management, progressive analysis and improvement, and challenge administration experience.
More details about USRA is obtainable at www.usra.edu.
‘;
Source link
