{"id":20207,"date":"2026-09-29T09:49:07","date_gmt":"2026-09-29T13:49:07","guid":{"rendered":"https:\/\/wp.glbgpt.com\/?p=20207"},"modified":"2026-09-29T09:49:08","modified_gmt":"2026-09-29T13:49:08","slug":"nasa-ibm-lunar-foundation-model","status":"publish","type":"post","link":"https:\/\/wp.glbgpt.com\/nl\/hub\/nasa-ibm-lunar-foundation-model","title":{"rendered":"NASA-IBM Lunar Foundation Model: toepassingen, resultaten en toegang"},"content":{"rendered":"<p class=\"wp-block-paragraph\"><strong>The NASA-IBM Lunar Foundation Model is an open AI model for analyzing lunar observations.<\/strong> Announced on September 10, 2026, it can be adapted to detect craters, outline volcanic features, and predict patterns in an ice-prospectivity map. Its reported gains concern specific research benchmarks: the headline ice result is lower prediction error, not the discovery of 22% more water on the Moon.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">For students and researchers, the useful starting points are the <a href=\"https:\/\/huggingface.co\/nasa-ibm-ai4science\/NASA-IBM-Lunar-Foundation-Model\">offici\u00eble modelkaart<\/a>, the technical report, and the released code. Together they explain what the model learned, which tasks were evaluated, and what you would need to build your own experiment.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">If your first job is understanding the research, <a href=\"https:\/\/www.glbgpt.com\/home?inviter=hub_content_home&amp;login=1\">GlobalGPT<\/a> brings source discovery, document analysis, reasoning, and writing into one multi-model workspace. It offers an affordable way to move from unfamiliar scientific terms to a clear research brief through a single subscription.<\/p>\n\n\n\n<div class=\"wp-block-group is-layout-constrained wp-block-group-is-layout-constrained\">\n<figure class=\"wp-block-image size-large\"><img fetchpriority=\"high\" decoding=\"async\" width=\"1024\" height=\"645\" src=\"https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/07\/glb-1024x645.png\" alt=\"GlobalGPT-dashboard\" class=\"wp-image-19280\" srcset=\"https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/07\/glb-1024x645.png 1024w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/07\/glb-300x189.png 300w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/07\/glb-18x11.png 18w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/07\/glb-767x483.png 767w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/07\/glb-1536x967.png 1536w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/07\/glb-2048x1289.png 2048w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<div class=\"wp-block-buttons is-content-justification-center is-layout-flex wp-container-core-buttons-is-layout-3e41869c wp-block-buttons-is-layout-flex\">\n<div class=\"wp-block-button\"><a class=\"wp-block-button__link has-black-color has-text-color has-background has-link-color wp-element-button\" href=\"https:\/\/www.glbgpt.com\/home?inviter=hub_popup&amp;login=1\" style=\"background:linear-gradient(135deg,rgba(250,183,0,0.55) 0%,rgba(255,106,0,0.66) 96%)\"><strong>Probeer meer dan 100 topmodellen op GlobalGPT<\/strong><\/a><\/div>\n<\/div>\n<\/div>\n\n\n\n<nav class=\"glb-lunar-toc\" aria-label=\"Inhoudsopgave\" style=\"box-sizing:border-box;margin:32px 0;padding:22px;border:1px solid #b9cbd7;border-radius:6px;background:#f4f8fb;color:#173b4b;max-width:100%;overflow-wrap:anywhere;font-family:system-ui,-apple-system,BlinkMacSystemFont,&quot;Segoe UI&quot;,sans-serif;line-height:1.65\"><p class=\"glb-lunar-toc-title\" style=\"box-sizing:border-box;margin:0 0 14px;font-size:1.05rem;font-weight:750;color:#173b4b\">Explore the lunar model<\/p><ol style=\"box-sizing:border-box;display:grid;grid-template-columns:repeat(auto-fit,minmax(min(100%,220px),1fr));gap:10px 30px;padding-left:23px;margin:0\"><li style=\"box-sizing:border-box;margin:0;line-height:1.55;color:#274756\"><a href=\"#what-is-it\" style=\"box-sizing:border-box;color:#12516a;text-decoration:underline;text-underline-offset:3px;overflow-wrap:anywhere\">What is the NASA-IBM Lunar Foundation Model?<\/a><\/li><li style=\"box-sizing:border-box;margin:0;line-height:1.55;color:#274756\"><a href=\"#data-and-design\" style=\"box-sizing:border-box;color:#12516a;text-decoration:underline;text-underline-offset:3px;overflow-wrap:anywhere\">What data does it learn from?<\/a><\/li><li style=\"box-sizing:border-box;margin:0;line-height:1.55;color:#274756\"><a href=\"#what-it-can-do\" style=\"box-sizing:border-box;color:#12516a;text-decoration:underline;text-underline-offset:3px;overflow-wrap:anywhere\">Three tasks it can support<\/a><\/li><li style=\"box-sizing:border-box;margin:0;line-height:1.55;color:#274756\"><a href=\"#benchmarks\" style=\"box-sizing:border-box;color:#12516a;text-decoration:underline;text-underline-offset:3px;overflow-wrap:anywhere\">What the benchmark results mean<\/a><\/li><li style=\"box-sizing:border-box;margin:0;line-height:1.55;color:#274756\"><a href=\"#access-and-cost\" style=\"box-sizing:border-box;color:#12516a;text-decoration:underline;text-underline-offset:3px;overflow-wrap:anywhere\">Where to get it and what access costs<\/a><\/li><li style=\"box-sizing:border-box;margin:0;line-height:1.55;color:#274756\"><a href=\"#research-workflow\" style=\"box-sizing:border-box;color:#12516a;text-decoration:underline;text-underline-offset:3px;overflow-wrap:anywhere\">Use AI to understand the research<\/a><\/li><li style=\"box-sizing:border-box;margin:0;line-height:1.55;color:#274756\"><a href=\"#limitations\" style=\"box-sizing:border-box;color:#12516a;text-decoration:underline;text-underline-offset:3px;overflow-wrap:anywhere\">Limits that matter in practice<\/a><\/li><li style=\"box-sizing:border-box;margin:0;line-height:1.55;color:#274756\"><a href=\"#faq\" style=\"box-sizing:border-box;color:#12516a;text-decoration:underline;text-underline-offset:3px;overflow-wrap:anywhere\">Veelgestelde vragen<\/a><\/li><\/ol><\/nav>\n\n\n\n<h2 id=\"what-is-it\" class=\"wp-block-heading\">What is the NASA-IBM Lunar Foundation Model?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/science.nasa.gov\/science-research\/artificial-intelligence-lunar-foundation-model\/\">NASA\u2019s announcement<\/a> introduces the project as a foundation model for lunar science, developed with IBM. A foundation model learns reusable patterns from a broad collection of data; researchers then adapt those patterns to a narrower task. Here, the material is lunar remote-sensing data rather than everyday language.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The NASA-IBM LFM uses a vision-transformer encoder and decoder, with a ViT-B architecture. Its inputs include imagery and other observation products at different spatial resolutions. A scientist can adapt the resulting representation for detection, segmentation, or regression\u2014three ways of finding objects, outlining regions, and predicting numerical values.<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img decoding=\"async\" width=\"1440\" height=\"900\" src=\"https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/nasa-lunar-announcement-20260928_78ea7b61cc404657b5112e2a8bea709a.webp\" alt=\"NASA announcement introducing the NASA-IBM Lunar Foundation Model for lunar science.\" class=\"wp-image-20212\" srcset=\"https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/nasa-lunar-announcement-20260928_78ea7b61cc404657b5112e2a8bea709a.webp 1440w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/nasa-lunar-announcement-20260928_78ea7b61cc404657b5112e2a8bea709a-300x188.webp 300w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/nasa-lunar-announcement-20260928_78ea7b61cc404657b5112e2a8bea709a-1024x640.webp 1024w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/nasa-lunar-announcement-20260928_78ea7b61cc404657b5112e2a8bea709a-768x480.webp 768w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/nasa-lunar-announcement-20260928_78ea7b61cc404657b5112e2a8bea709a-18x12.webp 18w\" sizes=\"(max-width: 1440px) 100vw, 1440px\" \/><figcaption class=\"wp-element-caption\"><a href=\"https:\/\/science.nasa.gov\/science-research\/artificial-intelligence-lunar-foundation-model\/\">NASA announced the lunar model<\/a> on September 10, 2026, with public model and code access.<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">That makes the release useful to teams working with lunar maps and to students studying planetary machine learning. Reading the project is also worthwhile for anyone trying to understand how a scientific AI model is evaluated. Using its checkpoint, however, is a research-computing workflow rather than a chat conversation.<\/p>\n\n\n\n<h2 id=\"data-and-design\" class=\"wp-block-heading\">What data does it learn from?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">De <a href=\"https:\/\/huggingface.co\/nasa-ibm-ai4science\/NASA-IBM-Lunar-Foundation-Model\">model card<\/a> describes the SomBench training data: 963,609 Wide Angle Camera (WAC) tile bundles and 1,000,113 Narrow Angle Camera (NAC) bundles. The WAC scale is about 100 meters per pixel; the NAC scale is about 1 meter per pixel. This gives the model both broader terrain context and much finer surface detail, across roughly two million bundles.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/newsroom.ibm.com\/2026-09-10-ibm-and-nasa-release-open-source-ai-model-to-support-lunar-exploration\">IBM describes the wider data resource<\/a> as more than 30 aligned layers from nine instruments across four missions. The model card specifies 11 input modalities. These figures describe different stages of the data pipeline: source layers, packaged training examples, and types of input the model encodes.<\/p>\n\n\n\n<section class=\"glb-lunar-data\" aria-label=\"Lunar data and model inputs\" style=\"box-sizing:border-box;margin:28px 0;padding:24px;background:#f1f6f5;border:1px solid #bad0cb;border-radius:6px;color:#183d38;max-width:100%;overflow-wrap:anywhere;font-family:system-ui,-apple-system,BlinkMacSystemFont,&quot;Segoe UI&quot;,sans-serif;line-height:1.65\"><p class=\"glb-lunar-data-title\" style=\"box-sizing:border-box;color:#173e37;line-height:1.6;margin:0 0 18px;font-weight:750;font-size:1.2rem\">One project, three different counting units<\/p><div class=\"glb-lunar-data-grid\" style=\"box-sizing:border-box;display:grid;grid-template-columns:repeat(auto-fit,minmax(min(100%,220px),1fr));gap:12px\"><div class=\"glb-lunar-data-step\" style=\"box-sizing:border-box;min-width:0;background:#fff;border:1px solid #c2d5cf;padding:16px;border-radius:4px\"><span class=\"glb-lunar-data-number\" style=\"box-sizing:border-box;display:block;color:#12594d;font-size:1.55rem;font-weight:750;margin-bottom:8px\">30+ source layers<\/span><p style=\"box-sizing:border-box;color:#294e45;line-height:1.6;margin:8px 0\">IBM describes a data resource assembled from nine instruments across four missions.<\/p><\/div><div class=\"glb-lunar-data-step\" style=\"box-sizing:border-box;min-width:0;background:#fff;border:1px solid #c2d5cf;padding:16px;border-radius:4px\"><span class=\"glb-lunar-data-number\" style=\"box-sizing:border-box;display:block;color:#12594d;font-size:1.55rem;font-weight:750;margin-bottom:8px\">1,963,722 bundles<\/span><p style=\"box-sizing:border-box;color:#294e45;line-height:1.6;margin:8px 0\">963,609 WAC bundles and 1,000,113 NAC bundles provide two scales of training material.<\/p><\/div><div class=\"glb-lunar-data-step\" style=\"box-sizing:border-box;min-width:0;background:#fff;border:1px solid #c2d5cf;padding:16px;border-radius:4px\"><span class=\"glb-lunar-data-number\" style=\"box-sizing:border-box;display:block;color:#12594d;font-size:1.55rem;font-weight:750;margin-bottom:8px\">11 input modalities<\/span><p style=\"box-sizing:border-box;color:#294e45;line-height:1.6;margin:8px 0\">The model encodes different kinds of observations, including imagery and terrain products.<\/p><\/div><\/div><p class=\"glb-lunar-data-note\" style=\"box-sizing:border-box;color:#35584f;line-height:1.6;margin:16px 0 0;font-size:.9rem\">Source layers, tile bundles, and input modalities count different things. Sources: <a href=\"https:\/\/newsroom.ibm.com\/2026-09-10-ibm-and-nasa-release-open-source-ai-model-to-support-lunar-exploration\" style=\"box-sizing:border-box;color:#10594e;text-decoration:underline\">IBM\u2019s dataset description<\/a> en de <a href=\"https:\/\/huggingface.co\/nasa-ibm-ai4science\/NASA-IBM-Lunar-Foundation-Model\" style=\"box-sizing:border-box;color:#10594e;text-decoration:underline\">offici\u00eble modelkaart<\/a>.<\/p><\/section>\n\n\n\n<p class=\"wp-block-paragraph\">Lighting is part of the design, too. The model receives observation geometry, including illumination angles, as explicit context. This helps distinguish how a surface looks under particular lighting from the underlying terrain patterns it is learning.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">A pixel representing roughly 100 meters of terrain cannot reveal the same details as a pixel representing roughly 1 meter. That is why the WAC and NAC crater results deserve separate attention. A gain on broad-scale imagery does not establish the same gain on small surface features.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The project follows the TerraMind pretraining approach but trains on lunar data from scratch. It should therefore be understood as a model built for lunar observations, rather than an Earth-observation checkpoint with a new name. The relevant question is how its learned representation helps a particular lunar task.<\/p>\n\n\n\n<h2 id=\"what-it-can-do\" class=\"wp-block-heading\">Three tasks it can support<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The published evaluations cover three task families. Each produces a different kind of answer, so the meaning of \u201cbetter\u201d changes with the task.<\/p>\n\n\n\n<div class=\"glb-lunar-table\" style=\"max-width:100%;overflow-x:auto;margin:28px 0;border:1px solid #c9d6df;border-radius:4px;box-sizing:border-box;overflow-wrap:anywhere;font-family:system-ui,-apple-system,BlinkMacSystemFont,&quot;Segoe UI&quot;,sans-serif;line-height:1.65\"><table style=\"border-collapse:collapse;width:100%;min-width:590px;font-size:.94rem;line-height:1.6\"><thead><tr><th scope=\"col\" style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#edf3f6;color:#173f51\">Taak<\/th><th scope=\"col\" style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#edf3f6;color:#173f51\">Uitgang<\/th><th scope=\"col\" style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#edf3f6;color:#173f51\">What the result can support<\/th><\/tr><\/thead><tbody><tr><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">Crater detection<\/td><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">Bounding boxes around craters in WAC or NAC imagery<\/td><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">Comparing how well a detector finds and locates labeled craters.<\/td><\/tr><tr><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">Irregular Mare Patch segmentation<\/td><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">A pixel-level outline of volcanic surface features<\/td><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">Studying how closely predicted regions match labeled feature boundaries.<\/td><\/tr><tr><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">Polar ice-prospectivity regression<\/td><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">A continuous prediction of a knowledge-driven prospectivity map<\/td><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">Studying agreement with a modeled target for possible ice-related conditions.<\/td><\/tr><\/tbody><\/table><\/div>\n\n\n\n<p class=\"wp-block-paragraph\">Crater detection and volcanic-feature segmentation can help researchers organize and analyze imagery. Their usefulness depends on the labels, scale, and evaluation area. A model that performs well on a defined test set still needs assessment on the data and scientific question of a new project.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Ice prospectivity needs the most careful interpretation.<\/strong> The target is a knowledge-driven map of where ice may be plausible, not a direct measurement of ice deposits. Better agreement with that map can support research into patterns and candidate regions. It does not tell us that an additional quantity of water has been found.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img decoding=\"async\" width=\"760\" height=\"500\" src=\"https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-model-intended-use-20260928_45c0828921cb4df7a9802ad5b4edb59a.webp\" alt=\"Official lunar model card describing research uses and limits of ice-prospectivity predictions.\" class=\"wp-image-20213\" srcset=\"https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-model-intended-use-20260928_45c0828921cb4df7a9802ad5b4edb59a.webp 760w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-model-intended-use-20260928_45c0828921cb4df7a9802ad5b4edb59a-300x197.webp 300w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-model-intended-use-20260928_45c0828921cb4df7a9802ad5b4edb59a-18x12.webp 18w\" sizes=\"(max-width: 760px) 100vw, 760px\" \/><figcaption class=\"wp-element-caption\">De <a href=\"https:\/\/huggingface.co\/nasa-ibm-ai4science\/NASA-IBM-Lunar-Foundation-Model\">offici\u00eble modelkaart<\/a> defines ice prospectivity as a modeled target and rules out landing-site certification.<\/figcaption><\/figure>\n\n\n\n<h2 id=\"benchmarks\" class=\"wp-block-heading\">What the benchmark results mean<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">De <a href=\"https:\/\/huggingface.co\/nasa-ibm-ai4science\/NASA-IBM-Lunar-Foundation-Model\/blob\/main\/NI_LFM_Technical_Report.pdf\">technical report<\/a> provides means and standard deviations across five random seeds. Its results compare adapted lunar-model configurations with named vision-model baselines. We checked the published tables; we did not reproduce the lunar experiments.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The metrics answer different questions. Mean average precision (mAP) evaluates object detections; intersection over union (IoU) measures overlap between predicted and labeled regions; root mean squared error (RMSE) measures numerical prediction error. Higher mAP and IoU are better, while lower RMSE is better. They should not be combined into one accuracy score.<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"850\" height=\"680\" src=\"https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-model-benchmarks-20260928_ce8b39a4f9d94857b7917f4ff77d7e9f.webp\" alt=\"Official benchmark table comparing the lunar model with ImageNet-pretrained baselines.\" class=\"wp-image-20211\" srcset=\"https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-model-benchmarks-20260928_ce8b39a4f9d94857b7917f4ff77d7e9f.webp 850w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-model-benchmarks-20260928_ce8b39a4f9d94857b7917f4ff77d7e9f-300x240.webp 300w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-model-benchmarks-20260928_ce8b39a4f9d94857b7917f4ff77d7e9f-768x614.webp 768w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-model-benchmarks-20260928_ce8b39a4f9d94857b7917f4ff77d7e9f-15x12.webp 15w\" sizes=\"(max-width: 850px) 100vw, 850px\" \/><figcaption class=\"wp-element-caption\">De <a href=\"https:\/\/huggingface.co\/nasa-ibm-ai4science\/NASA-IBM-Lunar-Foundation-Model\">authors\u2019 benchmark table<\/a> reports an ice-prospectivity gain and comparable meter-scale crater and volcanic-feature results.<\/figcaption><\/figure>\n\n\n\n<section class=\"glb-lunar-results\" aria-label=\"Author-reported lunar-model benchmarks\" style=\"box-sizing:border-box;margin:30px 0;padding:24px;background:#102c3b;border:1px solid #244f61;border-radius:6px;color:#f0f7fa;font:inherit;max-width:100%;overflow-wrap:anywhere;font-family:system-ui,-apple-system,BlinkMacSystemFont,&quot;Segoe UI&quot;,sans-serif;line-height:1.65\"><p class=\"glb-lunar-results-title\" style=\"box-sizing:border-box;color:#fff;line-height:1.35;margin:0 0 9px;font-size:1.3rem;font-weight:750\">Ice-prospectivity error: the clearest reported gain<\/p><p style=\"box-sizing:border-box;color:#e0edf3;line-height:1.65;margin:12px 0\">RMSE fell from <strong style=\"box-sizing:border-box;color:#fff\">0.0377 to 0.0293<\/strong> against SwinV2-B, a <strong style=\"box-sizing:border-box;color:#fff\">22.3% relative reduction<\/strong>. Lower RMSE means closer agreement with the benchmark\u2019s ice-prospectivity map.<\/p><div class=\"glb-lunar-results-bars\" style=\"box-sizing:border-box;margin:22px 0\"><div class=\"glb-lunar-results-label\" style=\"box-sizing:border-box;display:flex;justify-content:space-between;gap:12px;color:#e9f5f9;font-size:.94rem;margin:15px 0 7px\"><span style=\"box-sizing:border-box\">SwinV2-B \u00b7 ImageNet pretrained<\/span><span style=\"box-sizing:border-box\">0.0377 \u00b1 0.0004<\/span><\/div><div class=\"glb-lunar-results-track\" aria-hidden=\"true\" style=\"box-sizing:border-box;height:15px;background:#264454;border-radius:2px\"><div class=\"glb-lunar-results-bar\" style=\"box-sizing:border-box;height:15px;background:#9bc9dd;border-radius:2px;width:100%\"><\/div><\/div><div class=\"glb-lunar-results-label\" style=\"box-sizing:border-box;display:flex;justify-content:space-between;gap:12px;color:#e9f5f9;font-size:.94rem;margin:15px 0 7px\"><span style=\"box-sizing:border-box\">NASA-IBM LFM \u00b7 full fine-tuning<\/span><span style=\"box-sizing:border-box\">0.0293 \u00b1 0.0013<\/span><\/div><div class=\"glb-lunar-results-track\" aria-hidden=\"true\" style=\"box-sizing:border-box;height:15px;background:#264454;border-radius:2px\"><div class=\"glb-lunar-results-bar glb-lunar-results-bar-lfm\" style=\"box-sizing:border-box;height:15px;background:#f5c56d;border-radius:2px;width:77.7188%\"><\/div><\/div><\/div><p class=\"glb-lunar-results-note\" style=\"box-sizing:border-box;color:#c7dce6;line-height:1.65;margin:12px 0;font-size:.89rem\">RMSE, lower is better. Bars start at zero. Mean \u00b1 standard deviation over five seeds. The target is a modeled prospectivity map, not a measurement of ice.<\/p><div class=\"glb-lunar-results-scroll\" role=\"region\" aria-label=\"Other lunar benchmark results\" tabindex=\"0\" style=\"box-sizing:border-box;max-width:100%;overflow-x:auto;margin-top:22px\"><table style=\"box-sizing:border-box;border-collapse:collapse;width:100%;min-width:650px;color:#e4eff4;background:#102c3b;font-size:.86rem\"><thead style=\"box-sizing:border-box\"><tr style=\"box-sizing:border-box\"><th style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#fff;background:#1d3d4d;line-height:1.5;font-weight:700;border:1px solid #c6d6d0\">Task and data<\/th><th style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#fff;background:#1d3d4d;line-height:1.5;font-weight:700;border:1px solid #c6d6d0\">Metric \u2191<\/th><th style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#fff;background:#1d3d4d;line-height:1.5;font-weight:700;border:1px solid #c6d6d0\">Lunar model<\/th><th style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#fff;background:#1d3d4d;line-height:1.5;font-weight:700;border:1px solid #c6d6d0\">Strong baseline<\/th><\/tr><\/thead><tbody style=\"box-sizing:border-box\"><tr style=\"box-sizing:border-box\"><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">WAC craters \u00b7 50% training data<\/td><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">mAP<\/td><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">0.2541 \u00b1 0.0018<br style=\"box-sizing:border-box\">full fine-tuning<\/td><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">0.2313 \u00b1 0.0027<br style=\"box-sizing:border-box\">SwinV2-B<\/td><\/tr><tr style=\"box-sizing:border-box\"><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">WAC craters \u00b7 100% training data<\/td><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">mAP<\/td><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">0.2581 \u00b1 0.0017<br style=\"box-sizing:border-box\">LoRA<\/td><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">0.2420 \u00b1 0.0047<br style=\"box-sizing:border-box\">SwinV2-B<\/td><\/tr><tr style=\"box-sizing:border-box\"><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">NAC craters \u00b7 100% training data<\/td><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">mAP<\/td><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">0.1543 \u00b1 0.0098<br style=\"box-sizing:border-box\">LoRA<\/td><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">0.1552 \u00b1 0.0086<br style=\"box-sizing:border-box\">SwinV2-B<\/td><\/tr><tr style=\"box-sizing:border-box\"><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">Irregular Mare Patches<\/td><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">IoU1<\/td><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">0.5709 \u00b1 0.0114<br style=\"box-sizing:border-box\">frozen encoder<\/td><td style=\"box-sizing:border-box;padding:11px 10px;border-bottom:1px solid #38586a;text-align:left;vertical-align:top;color:#e4eff4;background:#102c3b;line-height:1.5;border:1px solid #c6d6d0\">0.5687 \u00b1 0.0181<br style=\"box-sizing:border-box\">ConvNeXtV2-B<\/td><\/tr><\/tbody><\/table><\/div><p style=\"box-sizing:border-box;color:#e0edf3;line-height:1.65;margin:12px 0\">The authors describe meter-scale crater and volcanic-feature results as comparable to strong baselines. Their small differences should be read alongside the run-to-run variation.<\/p><p class=\"glb-lunar-results-note\" style=\"box-sizing:border-box;color:#c7dce6;line-height:1.65;margin:12px 0;font-size:.89rem\">Bron: <a href=\"https:\/\/huggingface.co\/nasa-ibm-ai4science\/NASA-IBM-Lunar-Foundation-Model\/blob\/main\/NI_LFM_Technical_Report.pdf\" style=\"box-sizing:border-box;color:#94dbf2;text-decoration:underline\">NASA-IBM technical report<\/a>, Tables 4\u20137, pages 16 and 18. These are the authors\u2019 results, not an independent reproduction.<\/p><\/section>\n\n\n\n<h3 class=\"wp-block-heading\">The 22% ice result is a reduction in error<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">For ice prospectivity, the strongest reported lunar configuration has RMSE of 0.0293, versus 0.0377 for SwinV2-B. The calculation is <strong>(0.0377 \u2212 0.0293) \u00f7 0.0377 \u00d7 100 = 22.3%<\/strong>. RMSE measures prediction error against the target map; lower is better. This percentage is neither an accuracy score nor a change in the amount of lunar ice.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Crater results depend on the metric and data scale<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">With 50% of the WAC training data, the lunar model records mAP of 0.2541 versus 0.2313 for SwinV2-B. That same lunar-model result also exceeds the baseline\u2019s 0.2420 with 100% of the training data. The first comparison uses equal data fractions; the second is evidence of label efficiency in this particular experiment.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">IBM\u2019s roughly 19% crater improvement can be reconciled with the report\u2019s AP@75 values at 50% training data: 0.2213 versus 0.1862, an 18.9% relative increase. AP@75 evaluates detections at a particular overlap threshold. It should not be relabeled as a 19% increase in the table\u2019s mAP metric.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">At the finer NAC scale, the lunar model\u2019s mAP is 0.1543 and the baseline\u2019s is 0.1552. For Irregular Mare Patches, the reported foreground IoU values are 0.5709 and 0.5687 against ConvNeXtV2-B. The authors describe these tasks as comparable to strong baselines. Small differences, alongside the reported variation between runs, do not support a claim that the lunar model wins everywhere.<\/p>\n\n\n\n<h2 id=\"access-and-cost\" class=\"wp-block-heading\">Where to get it and what access costs<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Start from the official resources below. The model weights and code carry an Apache-2.0 license, and the checkpoint is publicly available. The repository specifically describes a <strong>fine-tuning and inference release<\/strong>; it explicitly excludes pretraining code.<\/p>\n\n\n\n<section class=\"glb-lunar-access\" aria-label=\"Official lunar model resources\" style=\"box-sizing:border-box;margin:28px 0;border:1px solid #b9cbd7;padding:24px;border-radius:6px;background:#f8fafc;color:#173544;max-width:100%;overflow-wrap:anywhere;font-family:system-ui,-apple-system,BlinkMacSystemFont,&quot;Segoe UI&quot;,sans-serif;line-height:1.65\"><p class=\"glb-lunar-access-title\" style=\"box-sizing:border-box;color:#173544;line-height:1.65;margin:0 0 16px;font-weight:750;font-size:1.2rem\">Choose the resource for your next step<\/p><div class=\"glb-lunar-access-grid\" style=\"box-sizing:border-box;display:grid;grid-template-columns:repeat(auto-fit,minmax(min(100%,220px),1fr));gap:12px\"><div class=\"glb-lunar-access-item\" style=\"box-sizing:border-box;padding:16px;min-width:0;border:1px solid #ced9df;border-radius:4px;background:#fff\"><a href=\"https:\/\/huggingface.co\/nasa-ibm-ai4science\/NASA-IBM-Lunar-Foundation-Model\" style=\"box-sizing:border-box;color:#12546c;font-weight:700;text-decoration:underline;overflow-wrap:anywhere\">Read the model card<\/a><p style=\"box-sizing:border-box;color:#2c4958;line-height:1.65;margin:9px 0 0\">Start with intended uses, input data, licensing, and the limitations.<\/p><\/div><div class=\"glb-lunar-access-item\" style=\"box-sizing:border-box;padding:16px;min-width:0;border:1px solid #ced9df;border-radius:4px;background:#fff\"><a href=\"https:\/\/huggingface.co\/nasa-ibm-ai4science\/NASA-IBM-Lunar-Foundation-Model\/tree\/main\" style=\"box-sizing:border-box;color:#12546c;font-weight:700;text-decoration:underline;overflow-wrap:anywhere\">Download weights and configuration<\/a><p style=\"box-sizing:border-box;color:#2c4958;line-height:1.65;margin:9px 0 0\">The Files tab contains the released checkpoint and accompanying files.<\/p><\/div><div class=\"glb-lunar-access-item\" style=\"box-sizing:border-box;padding:16px;min-width:0;border:1px solid #ced9df;border-radius:4px;background:#fff\"><a href=\"https:\/\/github.com\/NASA-IMPACT\/NASA-IBM-Lunar-Foundation-Model\" style=\"box-sizing:border-box;color:#12546c;font-weight:700;text-decoration:underline;overflow-wrap:anywhere\">Inspect fine-tuning and inference code<\/a><p style=\"box-sizing:border-box;color:#2c4958;line-height:1.65;margin:9px 0 0\">Use the README and TerraTorch task configurations to plan an experiment.<\/p><\/div><div class=\"glb-lunar-access-item\" style=\"box-sizing:border-box;padding:16px;min-width:0;border:1px solid #ced9df;border-radius:4px;background:#fff\"><a href=\"https:\/\/huggingface.co\/collections\/nasa-ibm-ai4science\/nasa-ibm-lunar-fm-and-downstream-models\" style=\"box-sizing:border-box;color:#12546c;font-weight:700;text-decoration:underline;overflow-wrap:anywhere\">Browse models and data resources<\/a><p style=\"box-sizing:border-box;color:#2c4958;line-height:1.65;margin:9px 0 0\">Follow the official collection to related downstream models and datasets.<\/p><\/div><div class=\"glb-lunar-access-item\" style=\"box-sizing:border-box;padding:16px;min-width:0;border:1px solid #ced9df;border-radius:4px;background:#fff\"><a href=\"https:\/\/huggingface.co\/nasa-ibm-ai4science\/NASA-IBM-Lunar-Foundation-Model\/blob\/main\/NI_LFM_Technical_Report.pdf\" style=\"box-sizing:border-box;color:#12546c;font-weight:700;text-decoration:underline;overflow-wrap:anywhere\">Read the technical report<\/a><p style=\"box-sizing:border-box;color:#2c4958;line-height:1.65;margin:9px 0 0\">Tables 4\u20137 provide the task results, settings, and variation across runs.<\/p><\/div><div class=\"glb-lunar-access-item\" style=\"box-sizing:border-box;padding:16px;min-width:0;border:1px solid #ced9df;border-radius:4px;background:#fff\"><a href=\"https:\/\/huggingface.co\/spaces\/nasa-ibm-ai4science\/nasa-ibm-lunar-foundation-model-generation-demo\" style=\"box-sizing:border-box;color:#12546c;font-weight:700;text-decoration:underline;overflow-wrap:anywhere\">Visit the official demonstration<\/a><p style=\"box-sizing:border-box;color:#2c4958;line-height:1.65;margin:9px 0 0\">Explore the project\u2019s demo entry. We did not execute a generation in this review.<\/p><\/div><\/div><p class=\"glb-lunar-access-note\" style=\"box-sizing:border-box;color:#2c4958;line-height:1.65;margin:9px 0 0;font-size:.9rem;margin-top:16px\">The model card listed no deployed Hugging Face Inference Provider when checked on September 28, 2026. A downloadable checkpoint and a demonstration page are separate access routes.<\/p><\/section>\n\n\n\n<p class=\"wp-block-paragraph\">If you want to adapt the model, choose the downstream task before setting up the experiment. The repository supplies TerraTorch-compatible components and configurations that connect the checkpoint, task, and data. TerraTorch is the software framework used to organize that training and evaluation workflow.<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>Read the model card and the README, then choose detection, segmentation, or regression based on the scientific question.<\/li>\n\n\n\n<li>Obtain the corresponding data and official checkpoint, and inspect the repository\u2019s task configuration and environment instructions.<\/li>\n\n\n\n<li>Prepare the data for that configuration and run the relevant fine-tuning or inference workflow.<\/li>\n\n\n\n<li>Evaluate against an appropriate baseline and retain the data split, settings, and outputs so the result can be checked.<\/li>\n<\/ol>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1440\" height=\"900\" src=\"https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-github-access-20260928_0ed29da82aa24945999f774b4713ac5d.webp\" alt=\"Official GitHub README describing the lunar model fine-tuning and inference release.\" class=\"wp-image-20210\" srcset=\"https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-github-access-20260928_0ed29da82aa24945999f774b4713ac5d.webp 1440w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-github-access-20260928_0ed29da82aa24945999f774b4713ac5d-300x188.webp 300w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-github-access-20260928_0ed29da82aa24945999f774b4713ac5d-1024x640.webp 1024w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-github-access-20260928_0ed29da82aa24945999f774b4713ac5d-768x480.webp 768w, https:\/\/wp.glbgpt.com\/wp-content\/uploads\/2026\/09\/lunar-github-access-20260928_0ed29da82aa24945999f774b4713ac5d-18x12.webp 18w\" sizes=\"(max-width: 1440px) 100vw, 1440px\" \/><figcaption class=\"wp-element-caption\">De <a href=\"https:\/\/github.com\/NASA-IMPACT\/NASA-IBM-Lunar-Foundation-Model\">offici\u00eble repository<\/a> provides fine-tuning and inference code; it does not include pretraining code.<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">Open access removes a model-license purchase from the starting process, but running an experiment still has costs. You need compute, storage, data preparation, and time for evaluation. A released checkpoint does not automatically come with a hosted API or free GPU time, and the sources do not establish a universal minimum GPU for every downstream task.<\/p>\n\n\n\n<div class=\"glb-lunar-table\" style=\"max-width:100%;overflow-x:auto;margin:28px 0;border:1px solid #c9d6df;border-radius:4px;box-sizing:border-box;overflow-wrap:anywhere;font-family:system-ui,-apple-system,BlinkMacSystemFont,&quot;Segoe UI&quot;,sans-serif;line-height:1.65\"><table style=\"border-collapse:collapse;width:100%;min-width:590px;font-size:.94rem;line-height:1.6\"><thead><tr><th scope=\"col\" style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#edf3f6;color:#173f51\">Je doel<\/th><th scope=\"col\" style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#edf3f6;color:#173f51\">Resource or cost to plan for<\/th><\/tr><\/thead><tbody><tr><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">Read and assess the release<\/td><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">Public announcements, model card, report, and repository.<\/td><\/tr><tr><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">Run or adapt the scientific model<\/td><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">Suitable compute and storage, relevant data, environment setup, and evaluation work.<\/td><\/tr><tr><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">Understand papers and write a research brief<\/td><td style=\"padding:13px 15px;text-align:left;vertical-align:top;border:1px solid #d5dfe5;background:#fff;color:#263c48\">A research assistant workspace such as GlobalGPT, with model access and document tools in one subscription.<\/td><\/tr><\/tbody><\/table><\/div>\n\n\n\n<h2 id=\"research-workflow\" class=\"wp-block-heading\">Use AI to understand the research<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">A productive first question is narrow: \u201cWhat does the reported 22% improvement actually measure?\u201d Answering it requires a target, a metric, a baseline, and a calculation. Asking for those four items is more useful than asking an assistant whether the model is impressive.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">GlobalGPT lets you carry that reading work through source discovery, document explanation, and writing without moving between separate tools for every step. For a repeatable method, <a href=\"https:\/\/www.glbgpt.com\/hub\/how-to-use-perplexity-for-research\/\">build a source-checked research brief<\/a> with a short source list and explicit questions before requesting a polished summary.<\/p>\n\n\n\n<section class=\"glb-lunar-workflow\" aria-label=\"A three-step research workflow\" style=\"box-sizing:border-box;margin:28px 0;padding:24px;border:1px solid #c2d2de;border-radius:6px;background:#eef5fa;color:#183c51;max-width:100%;overflow-wrap:anywhere;font-family:system-ui,-apple-system,BlinkMacSystemFont,&quot;Segoe UI&quot;,sans-serif;line-height:1.65\"><p class=\"glb-lunar-workflow-title\" style=\"box-sizing:border-box;font-size:1.2rem;font-weight:750;margin:0 0 16px;color:#183c51\">From official sources to a useful briefing<\/p><ol style=\"box-sizing:border-box;margin:0;padding-left:24px\"><li style=\"box-sizing:border-box;padding:0 0 14px 4px;color:#24495f;line-height:1.65\"><strong style=\"box-sizing:border-box;color:#153e55\">Collect the primary material.<\/strong> Start with NASA\u2019s announcement, then add the model card, technical report, and README. You can <a href=\"https:\/\/www.glbgpt.com\/perplexity?inviter=hub_content_perplexity&amp;login=1\" style=\"box-sizing:border-box;color:#125977;text-decoration:underline\">research with Perplexity on GlobalGPT<\/a> to locate supporting sources.<\/li><li style=\"box-sizing:border-box;padding:0 0 14px 4px;color:#24495f;line-height:1.65\"><strong style=\"box-sizing:border-box;color:#153e55\">Ask focused questions.<\/strong> Have an assistant explain mAP, IoU, and RMSE separately, then list the target, baseline, and evaluation setting behind each claim.<\/li><li style=\"box-sizing:border-box;padding:0 0 14px 4px;color:#24495f;line-height:1.65;padding-bottom:0\"><strong style=\"box-sizing:border-box;color:#153e55\">Check and write.<\/strong> Reopen the cited table, verify each number, and turn the checked notes into a briefing that preserves the scientific limits.<\/li><\/ol><\/section>\n\n\n\n<p class=\"wp-block-paragraph\">We tested a small version of this workflow with GPT-5.6 Sol. It received a packet distilled from the four official sources and had to produce a short briefing, classify six claims, calculate the RMSE reduction, and suggest practical next steps. The task used supplied text; it did not browse for new evidence.<\/p>\n\n\n\n<section class=\"glb-lunar-setup\" aria-label=\"Research briefing test setup\" style=\"box-sizing:border-box;margin:28px 0;padding:22px;border:1px solid #b9cbd7;border-left:4px solid #13576c;border-radius:6px;background:#f3f8fb;color:#172f3b;font:inherit;max-width:100%;overflow-wrap:anywhere;font-family:system-ui,-apple-system,BlinkMacSystemFont,&quot;Segoe UI&quot;,sans-serif;line-height:1.65\"><p class=\"glb-lunar-setup-title\" style=\"box-sizing:border-box;margin:0 0 10px;color:#123443;line-height:1.65;font-size:1.18rem;font-weight:700\">A source-based briefing test<\/p><p style=\"box-sizing:border-box;margin:8px 0;color:#243f4c;line-height:1.65\">We gave GPT-5.6 Sol a short packet drawn from NASA\u2019s announcement, the official model card, the technical report, and the GitHub README. It had to write a briefing, check six claims, and explain the RMSE calculation.<\/p><dl style=\"box-sizing:border-box;display:block;grid-template-columns:repeat(auto-fit,minmax(min(100%,220px),1fr));gap:12px 24px;margin:16px 0 0\"><div style=\"box-sizing:border-box\"><dt style=\"box-sizing:border-box;font-weight:700;color:#234453\">Workspace and date<\/dt><dd style=\"box-sizing:border-box;margin:4px 0 14px;color:#2c4855\">GlobalGPT \u00b7 September 28, 2026<\/dd><\/div><div style=\"box-sizing:border-box\"><dt style=\"box-sizing:border-box;font-weight:700;color:#234453\">Beleid uitvoeren<\/dt><dd style=\"box-sizing:border-box;margin:4px 0 14px;color:#2c4855\">One task; first complete answer; no retries<\/dd><\/div><div style=\"box-sizing:border-box\"><dt style=\"box-sizing:border-box;font-weight:700;color:#234453\">Invoer<\/dt><dd style=\"box-sizing:border-box;margin:4px 0 14px;color:#2c4855\">355-word source packet with four source IDs<\/dd><\/div><div style=\"box-sizing:border-box\"><dt style=\"box-sizing:border-box;font-weight:700;color:#234453\">Toepassingsgebied<\/dt><dd style=\"box-sizing:border-box;margin:4px 0 14px;color:#2c4855\">Reading and synthesis; no live search or lunar-model execution<\/dd><\/div><div style=\"box-sizing:border-box\"><dt style=\"box-sizing:border-box;font-weight:700;color:#234453\">Response settings<\/dt><dd style=\"box-sizing:border-box;margin:4px 0 14px;color:#2c4855\">Output-token limit: 4,096; all other sampling controls left at service defaults<\/dd><\/div><div style=\"box-sizing:border-box\"><dt style=\"box-sizing:border-box;font-weight:700;color:#234453\">Usage and cost<\/dt><dd style=\"box-sizing:border-box;margin:4px 0 14px;color:#2c4855\">2,154 total tokens reported; the response did not include a monetary charge<\/dd><\/div><\/dl><\/section>\n\n\n\n<section class=\"glb-lunar-t01\" aria-label=\"Source-based briefing test result\" style=\"box-sizing:border-box;margin:28px 0;padding:24px;background:#f7fafc;border:1px solid #b9cbd7;border-radius:6px;color:#172f3b;font:inherit;max-width:100%;overflow-wrap:anywhere;font-family:system-ui,-apple-system,BlinkMacSystemFont,&quot;Segoe UI&quot;,sans-serif;line-height:1.65\"><p class=\"glb-lunar-t01-title\" style=\"box-sizing:border-box;color:#123443;line-height:1.35;margin:0 0 8px;font-size:1.3rem;font-weight:700\">The first answer kept the scientific claims in scope<\/p><div class=\"glb-lunar-t01-metrics\" style=\"box-sizing:border-box;display:grid;grid-template-columns:repeat(auto-fit,minmax(min(100%,220px),1fr));gap:10px;margin:18px 0\"><div class=\"glb-lunar-t01-metric\" style=\"box-sizing:border-box;padding:14px;background:#e3eef3;border:1px solid #c4d7df;border-radius:4px\"><span class=\"glb-lunar-t01-number\" style=\"box-sizing:border-box;display:block;font-size:1.55rem;font-weight:750;color:#123c4d;line-height:1.2\">6 \/ 6<\/span><span class=\"glb-lunar-t01-label\" style=\"box-sizing:border-box;display:block;font-size:.86rem;color:#294b5a;margin-top:6px;line-height:1.4\">Claim checks matched the supplied sources<\/span><\/div><div class=\"glb-lunar-t01-metric\" style=\"box-sizing:border-box;padding:14px;background:#e3eef3;border:1px solid #c4d7df;border-radius:4px\"><span class=\"glb-lunar-t01-number\" style=\"box-sizing:border-box;display:block;font-size:1.55rem;font-weight:750;color:#123c4d;line-height:1.2\">22.3%<\/span><span class=\"glb-lunar-t01-label\" style=\"box-sizing:border-box;display:block;font-size:.86rem;color:#294b5a;margin-top:6px;line-height:1.4\">Correct relative RMSE reduction<\/span><\/div><div class=\"glb-lunar-t01-metric\" style=\"box-sizing:border-box;padding:14px;background:#e3eef3;border:1px solid #c4d7df;border-radius:4px\"><span class=\"glb-lunar-t01-number\" style=\"box-sizing:border-box;display:block;font-size:1.55rem;font-weight:750;color:#123c4d;line-height:1.2\">14.0 s<\/span><span class=\"glb-lunar-t01-label\" style=\"box-sizing:border-box;display:block;font-size:.86rem;color:#294b5a;margin-top:6px;line-height:1.4\">Time for this complete response<\/span><\/div><\/div><p style=\"box-sizing:border-box;color:#25434f;line-height:1.65;margin:12px 0\">The answer produced a 112-word briefing, rejected the claim that the model had discovered 22% more real ice, and retained the restrictions on landing-site certification and pretraining-code access. Its three next steps pointed back to the official model card, checkpoint, and TerraTorch resources.<\/p><p class=\"glb-lunar-t01-note\" style=\"box-sizing:border-box;color:#405967;line-height:1.65;margin:12px 0;font-size:.91rem\">This was one reading task using supplied material. The source packet already stated the key limitations. The result does not establish performance on autonomous research or lunar remote-sensing inference.<\/p><details style=\"box-sizing:border-box;margin:14px 0 0;padding:12px 14px;background:#fff;border:1px solid #bed0dc;border-radius:4px\"><summary style=\"box-sizing:border-box;cursor:pointer;color:#174d62;font-weight:700;line-height:1.5\">View the exact prompt and source packet<\/summary><pre style=\"box-sizing:border-box;margin:14px 0 0;white-space:pre-wrap;overflow-wrap:anywhere;word-break:normal;max-width:100%;font:0.88rem\/1.65 ui-monospace,SFMono-Regular,Consolas,monospace;color:#203d4a;background:#fff\">SYSTEM\nYou are a careful research editor writing for a university student. Use only the supplied source packet. Do not browse. Distinguish reported benchmark measurements from real-world scientific measurements. Cite source IDs in square brackets. If a claim is not supported, say so plainly. Do not invent citations, prices, hardware requirements, or experimental results. Provide the requested answer, not hidden reasoning.\n\nUSER\nTurn the source packet into a practical research note. Return: (1) a plain-English briefing of 100\u2013130 words; (2) a six-row claim-check table using exactly Supported or Not supported, a source ID, and a one-sentence explanation for each supplied claim; (3) the relative RMSE reduction calculated from 0.0377 to 0.0293, rounded to one decimal place, with a formula and one sentence explaining what it does not measure; (4) three practical next steps for a student who wants to inspect or adapt the released model. Keep the whole response under 500 words.\n\nClaims to check:\n1. NASA and IBM publicly announced the model on September 10, 2026.\n2. The reported ice-prospectivity RMSE fell from 0.0377 for SwinV2-B to 0.0293 for the full-fine-tuned lunar model.\n3. The model discovered 22% more real water ice on the Moon.\n4. NASA has validated this model for landing-site certification and hazard clearance.\n5. The GitHub release includes the pretraining code.\n6. An open checkpoint guarantees that GPU compute and hosted inference are free.\n\nSource packet:\n[S1] NASA announcement, September 10, 2026.\nTitle: NASA, IBM Launch AI Foundation Model for Lunar Science.\nURL: https:\/\/science.nasa.gov\/science-research\/artificial-intelligence-lunar-foundation-model\/\nNASA announced a lunar-science foundation model developed with IBM. It is publicly hosted on Hugging Face, with code on GitHub. It can be adapted for crater mapping, volcanic-feature identification and polar ice research.\n\n[S2] Official NASA-IBM model card and technical report, checked September 28, 2026.\nModel card URL: https:\/\/huggingface.co\/nasa-ibm-ai4science\/NASA-IBM-Lunar-Foundation-Model\nTechnical report URL: https:\/\/huggingface.co\/nasa-ibm-ai4science\/NASA-IBM-Lunar-Foundation-Model\/blob\/main\/NI_LFM_Technical_Report.pdf\nThe report&#x27;s Table 7 (page 18) gives test-set root mean squared error (RMSE) for ice-prospectivity regression. Lower is better. Mean plus\/minus standard deviation over five random seeds: NASA-IBM LFM with full fine-tuning, 0.0293 +\/- 0.0013; ImageNet-pretrained SwinV2-B, 0.0377 +\/- 0.0004. The target is a knowledge-driven ice-prospectivity map. These are author-reported benchmark results, not an independent replication or a direct measurement of water ice.\nTable 4 (page 16), WAC craters with 50% training data: bounding-box mAP is 0.2541 +\/- 0.0018 for full-fine-tuned LFM and 0.2313 +\/- 0.0027 for SwinV2-B. In the same rows, AP@75 is 0.2213 +\/- 0.0028 and 0.1862 +\/- 0.0041 respectively. These metrics must not be interchanged.\n\n[S3] Official model card, Intended use \/ Out of scope \/ Limitations.\nURL: https:\/\/huggingface.co\/nasa-ibm-ai4science\/NASA-IBM-Lunar-Foundation-Model\nThe intended use is lunar remote sensing through fine-tuning or LoRA adaptation for detection, segmentation and dense regression. The model is not validated for landing-site certification or hazard clearance. Ice-prospectivity outputs regress a knowledge-driven map, not measured ice. Generated fields are qualitative probes, not calibrated scientific predictions. The model does not maintain an absolute geodetic reference frame. It has not been evaluated beyond the Moon. The listed license is Apache-2.0. The page reports no Inference Provider deployment. This packet provides no promise of free GPU compute or free hosted inference, and no hardware minimum or hosted API price.\n\n[S4] Official GitHub README, checked September 28, 2026.\nURL: https:\/\/github.com\/NASA-IMPACT\/NASA-IBM-Lunar-Foundation-Model\nThe repository describes a fine-tuning and inference release. Pretraining code is not included. Weights and configuration can be downloaded from the official Hugging Face model repository. TerraTorch-compatible datamodules, tasks, backbone wrappers and task configurations are provided. Fine-tuning and testing require the relevant data, checkpoint and task configuration. The README lists an Apache 2.0 license.\n<\/pre><\/details><details style=\"box-sizing:border-box;margin:14px 0 0;padding:12px 14px;background:#fff;border:1px solid #bed0dc;border-radius:4px\"><summary style=\"box-sizing:border-box;cursor:pointer;color:#174d62;font-weight:700;line-height:1.5\">View the complete first answer<\/summary><pre style=\"box-sizing:border-box;margin:14px 0 0;white-space:pre-wrap;overflow-wrap:anywhere;word-break:normal;max-width:100%;font:0.88rem\/1.65 ui-monospace,SFMono-Regular,Consolas,monospace;color:#203d4a;background:#fff\">## 1. Briefing\n\nNASA and IBM announced the Lunar Foundation Model on September 10, 2026. The public release provides weights on Hugging Face and fine-tuning and inference code on GitHub, targeting lunar remote-sensing tasks such as crater mapping, volcanic-feature identification and polar-ice research. Its reported ice-prospectivity benchmark improves RMSE over an ImageNet-pretrained SwinV2-B, but the target is a knowledge-driven map, not directly measured lunar ice. The result is author-reported across five random seeds and is not an independent replication. The model is intended for adaptation via full fine-tuning or LoRA, but is not validated for landing-site certification or hazard clearance. Pretraining code is absent, and the release promises neither free GPU compute nor hosted inference. [S1][S2][S3][S4]\n\n## 2. Claim-check table\n\n| Claim | Verdict | Source | Explanation |\n|---|---|---|---|\n| 1 | Supported | [S1] | NASA\u2019s announcement is dated September 10, 2026, and identifies the model as a NASA\u2013IBM project. |\n| 2 | Supported | [S2] | Table 7 reports RMSE values of 0.0377 for SwinV2-B and 0.0293 for the fully fine-tuned lunar model. |\n| 3 | Not supported | [S2] | The figures are benchmark errors against a knowledge-driven prospectivity map, not measurements of real lunar water ice. |\n| 4 | Not supported | [S3] | The model card explicitly says the model is not validated for landing-site certification or hazard clearance. |\n| 5 | Not supported | [S4] | The GitHub repository provides fine-tuning and inference resources but excludes pretraining code. |\n| 6 | Not supported | [S3] | The packet promises no free GPU compute or hosted inference, and reports no Inference Provider deployment. |\n\n## 3. Relative RMSE reduction\n\n\\[\n\\frac{0.0377-0.0293}{0.0377}\\times 100 = 22.3\\%\n\\]\n\nThis is a **22.3% relative reduction in author-reported benchmark RMSE**; it does not measure additional real water ice discovered or independently validated scientific performance. [S2]\n\n## 4. Practical next steps\n\n1. Read the model card\u2019s intended uses, limitations, and Apache-2.0 license before selecting a detection, segmentation, or regression task. [S3]\n2. Download the official checkpoint and configuration from Hugging Face, then inspect the GitHub backbone wrappers and task configurations. [S4]\n3. Obtain the relevant lunar dataset and use the provided TerraTorch-compatible components to fine-tune and test, documenting that pretraining code is unavailable and benchmark results require independent evaluation. [S2][S4]<\/pre><\/details><\/section>\n\n\n\n<p class=\"wp-block-paragraph\">The useful outcome was the preservation of the distinctions: benchmark error versus discovered ice, research use versus landing certification, and released adaptation code versus absent pretraining code. Because those limits were already in the source packet, this is evidence of a successful reading-and-synthesis task, not proof that an assistant would independently uncover every caveat.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">For a larger project, <a href=\"https:\/\/www.glbgpt.com\/hub\/how-to-use-chatgpt-deep-research-complete-tutorial-tips-and-best-practices\/\">plan a deeper AI-assisted research session<\/a> around separate questions about data, methods, evaluation, and limitations. Keep the source table beside the draft and verify every number again after editing. That habit matters more than how fluent the first summary sounds.<\/p>\n\n\n\n<h2 id=\"limitations\" class=\"wp-block-heading\">Limits that matter in practice<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The model card says the lunar model has not been validated for operational decisions such as landing-site certification or hazard clearance. A scientific model may help explore observations without being ready to certify that a spacecraft can safely land somewhere. Those uses require a different level of validation.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Generated fields also need caution. The model does not maintain an absolute geodetic reference frame, and generated values can drift from physical values. The authors describe generation as a qualitative probe rather than calibrated scientific prediction. A plausible-looking terrain or observation map should not be read as a measured product.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The checkpoint has not been evaluated beyond the Moon. Even within lunar research, the benchmark story is mixed: clear gains on some tasks and comparable results on others. The defensible lesson is to judge the model against the target, scale, data, and baseline that match your experiment.<\/p>\n\n\n\n<h2 id=\"faq\" class=\"wp-block-heading\">Veelgestelde vragen<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">Is the NASA-IBM Lunar Foundation Model free to use?<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">The official model card and code repository list an Apache-2.0 license, and the checkpoint is publicly available. Running experiments still requires suitable computing resources, data preparation, and task-specific evaluation. Open access to the model does not provide a guarantee of free GPU compute or hosted inference.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Is the NASA-IBM Lunar Foundation Model a chatbot?<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">It is a foundation model for lunar remote sensing, built to work with imagery, terrain products, and related observation data. Researchers adapt it for detection, segmentation, and regression tasks. A chat assistant can help explain the documentation, but interacting with a chatbot is a different workflow from running this model.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Did the model discover more water ice on the Moon?<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">The reported ice result measures how closely predictions match a knowledge-driven ice-prospectivity map. The best reported RMSE falls from 0.0377 to 0.0293 against the cited baseline, a relative reduction of about 22%. That is a benchmark improvement; it is not a measurement of additional ice discovered.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Can I explore the project without writing code?<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">You can read NASA\u2019s announcement, inspect the model card, explore the published benchmark tables, and visit the official demonstration page without building a training pipeline. Adapting the released checkpoint to a scientific task uses the repository\u2019s code, data, and TerraTorch configurations, followed by task-specific validation.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">How is the lunar model related to Prithvi and Surya?<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">They belong to the broader NASA\u2013IBM collaboration on foundation models for science. NASA describes Prithvi models for Earth observations and weather applications, Surya for heliophysics, and the lunar model for Moon observations. Their domains and intended tasks differ; the lunar checkpoint has not been evaluated beyond the Moon.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">The NASA-IBM Lunar Foundation Model offers a reusable starting point for lunar remote-sensing research. Use the official tables to understand its strengths, the model card to define its limits, and the released resources to plan a task-specific experiment. If your immediate goal is learning or writing, start by turning those sources into a checked briefing.<\/p>\n\n\n\n<section class=\"glb-lunar-cta\" aria-label=\"Build a research brief with GlobalGPT\" style=\"box-sizing:border-box;margin:34px 0 14px;padding:26px;border:1px solid #234e63;border-radius:6px;background:#123142;color:#eef8fc;max-width:100%;overflow-wrap:anywhere;font-family:system-ui,-apple-system,BlinkMacSystemFont,&quot;Segoe UI&quot;,sans-serif;line-height:1.65\"><p class=\"glb-lunar-cta-title\" style=\"box-sizing:border-box;color:#fff;line-height:1.4;margin:0 0 12px;font-size:1.35rem;font-weight:750\">Turn a technical paper into a briefing you can use<\/p><p style=\"box-sizing:border-box;color:#e2eef5;line-height:1.7;margin:12px 0 20px\">Bring your official sources to GlobalGPT, work through unfamiliar concepts, and shape the checked findings into clear notes or a finished draft\u2014all in one multi-model workspace.<\/p><a class=\"glb-lunar-cta-button\" href=\"https:\/\/www.glbgpt.com\/home?inviter=hub_content_home&amp;login=1\" style=\"box-sizing:border-box;display:inline-block;padding:12px 17px;border-radius:4px;background:#d9f2ef;color:#103b3b;font-weight:750;text-decoration:none;line-height:1.5;max-width:100%;overflow-wrap:anywhere\">Explore research with GlobalGPT<\/a><\/section>","protected":false},"excerpt":{"rendered":"<p>The NASA-IBM Lunar Foundation Model is an open AI model for analyzing lunar observations. Announced on September 10, 2026, it can be adapted to detect craters, outline volcanic features, and predict patterns in an ice-prospectivity map. Its reported gains concern specific research benchmarks: the headline ice result is lower prediction error, not the discovery of [&hellip;]<\/p>","protected":false},"author":7,"featured_media":20209,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"_seopress_robots_primary_cat":"","_seopress_titles_title":"NASA-IBM Lunar Foundation Model: Uses, Results & Access","_seopress_titles_desc":"Learn what the NASA-IBM Lunar Foundation Model does, how to read its ice and crater benchmarks, and where to get the open model, data, and research code.","_seopress_robots_index":"","footnotes":""},"categories":[7],"tags":[],"class_list":["post-20207","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-chat"],"acf":[],"_links":{"self":[{"href":"https:\/\/wp.glbgpt.com\/nl\/wp-json\/wp\/v2\/posts\/20207","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/wp.glbgpt.com\/nl\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/wp.glbgpt.com\/nl\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/wp.glbgpt.com\/nl\/wp-json\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/wp.glbgpt.com\/nl\/wp-json\/wp\/v2\/comments?post=20207"}],"version-history":[{"count":2,"href":"https:\/\/wp.glbgpt.com\/nl\/wp-json\/wp\/v2\/posts\/20207\/revisions"}],"predecessor-version":[{"id":20215,"href":"https:\/\/wp.glbgpt.com\/nl\/wp-json\/wp\/v2\/posts\/20207\/revisions\/20215"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/wp.glbgpt.com\/nl\/wp-json\/wp\/v2\/media\/20209"}],"wp:attachment":[{"href":"https:\/\/wp.glbgpt.com\/nl\/wp-json\/wp\/v2\/media?parent=20207"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/wp.glbgpt.com\/nl\/wp-json\/wp\/v2\/categories?post=20207"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/wp.glbgpt.com\/nl\/wp-json\/wp\/v2\/tags?post=20207"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}