NASA’S AI Foundation Models for Science: Current Initiatives, Workflow, Roadmap and Lessons Learned

NASA’s scientific data archives have grown rapidly, surpassing 150 petabytes and expected to exceed 500 petabytes by 2029. This expansion presents considerable challenges in data management and analysis. To address these, the Office of the Chief Science Data Officer (OCSDO) has implemented a strategy centered on AI foundation models (FMs) to enhance workflows across NASA’s five science divisions. These models, trained with self-supervised pretraining, create versatile, application-agnostic representations that can be adapted efficiently to specific tasks. The OCSDO’s ”5+1” strategy integrates tailored FMs for each science division with a large language model for cross-domain tasks. Notable initiatives include the INDUS language model suite covering all science divisions, the Prithvi-HLS model for optical remote sensing, and the Prithvi-WxC model for atmospheric analysis. These models reduce computational demands and data labeling needs while performing well on existing benchmarks. Current efforts focus on new models for heliophysics (SuryaSDO), lunar studies, and biological research, undertaken with new partnerships. The FM development process follows science standards prioritizing transparency, accuracy, and relevance. Key workflows encompass pretraining, adaptation, and inference.
Lessons learned from this work emphasize the value of balancing costs and benefits, fostering interdisciplinary collaboration, designing science use case-driven models, and maintaining rigorous validation. This paper details the design, implementation, and roadmap of NASA’s FMs, illustrating their potential role in advancing scientific discovery through AI-powered methodologies.