Supercomputing

NASA Foundation Models: Real Science, No Hallucinations

Image from animated foundation model flow chart
Foundation Models evolve through three phases: pre-training, fine-tuning, and inference. Pre-training requires large quantities of data and high-performance computing capacity. Fine-tuning requires nothing more than a four-node GPU cluster, and inference can be performed using a single GPU. Carlos Costa, IBM Research; Michael Little, NASA/Goddard

Foundation models (FM), and the more common large language models (LLM), have gained a lot of visibility as toys or high school paper generators. But the technology, used properly, can provide a far more effective tool for actual scientific insight, reusable by a number of scientific investigations into a single science domain. Accomplishing this requires careful curation of the training data and broad evaluation and validation of the model before new scientific investigations can be run.

NASA has five different experiments underway in preparing these models and testing and validating their output against conventional scientific data. If successful, these models can be used as if they are a new instrument, integrating the output of multiple instruments and enabling multi-domain investigations that are very labor intensive today, if possible at all. Examples include investigations of wildfire and air quality as well as integrating LLM and geospatial FM to enable conversation with the model to generate results. Two common problems plaguing confidence in generative FM must be solved before use in science. First, validation of results and trapping hallucinations before output. Second, an accompanying Uncertainty Quantification that qualifies the fidelity of any hyperrealistic output.

Quick Facts

Working together with the science discipline areas and NASA advanced supercomputing, we envision foundation models acting as virtual instruments with flexibility to support a wide range of experiments, democratizing remote sensing in NASA’s Year of Open Science.

Michael Little,
NASA Goddard Space Flight Center
  • The Prithvi-100M Foundation Model was pre-trained at IBM on Harmonized Landsat Sentinel-2 (HLS) data and is now preparing to train on GPUs at the NASA Advanced Supercomputing facility.
  • Downstream applications provide crop classification, flood maps, and burn scars to about 75% accuracy.
  • Inexpensive inference democratizes science.
  • Future builds include Prithvi-300M and additional foundation models in heliophysics and climate.

Researchers

  • Michael Little, NASA Goddard Space Flight Center
  • Manil Maskey, NASA Headquarters

More Information