John Ossorgin, Sandia National Laboratories; Forest Danford, Sandia National Laboratories; Kyle Merry, Sandia National Laboratories
Keywords: model zoo, hosting service, AI/ML, models, SDA, SSA, community, space control
Abstract:
The field of Space Domain Awareness (SDA) is critical for national defense, requiring rapid and accurate analysis of space-related data. To support the SDA community within the AI/ML field, we propose Cosmic Collaboratory, a neural network model hosting service specifically tailored for SDA applications. This platform, akin to Hugging Face and TorchHub, will allow researchers to share pre-trained, verified, and validated neural network models, facilitating collaboration and setting the gold standard for the development of new AI/ML SDA projects.
AI/ML models have demonstrated that they can contribute to SDA. At AMOS 2024, it became readily apparent that there was a community desire for a centralized place to share and discuss the latest models and advances and a need for a stronger community of AI/ML researchers in the SDA field. Due to the sparse nature of SDA data, many widely available, commodity/general pretrained models are not suited to the task of analyzing SDA data. SDA data tends to be black and white “blobs and streaks” signal to noise representations while most models are tuned more for image segmentation and classification tasks such as text recognition, or image recognition for medical purposes or autonomous driving.
Today, the average AI/ML researcher has access to over one million pretrained AI/ML models suited to almost any task imaginable, trained on many robust datasets, with clear paths of contact to the original developers of the models. However, in the SDA world, we cannot simply go to an open website and find a model that suits our needs. In fact, many of the preeminent pre-trained image classification models do not perform well on SDA data even with significant changes. This requires researchers to go back to basics with an untrained model, adjust its pipeline to process these data, and train it on SDA data specifically. This takes considerable valuable time and resources, which we may not always have. By providing SDA specific models to the wider SDA community, we would be able to enable quick turn around on many projects responding to new sensitive information. We would also foster a stronger SDA community as a whole by promoting easy collaboration among researchers.
We want our models and datasets to be easily accessible by all members of the SDA community. Due to the potentially sensitive nature of our work, we do not want our pretrained models to be on the open web. We therefore propose to host the service on commodity platforms that many SDA members will already have access to, to ensure that only SDA community members have access to Cosmic Collaboratory and ensure the safety of our information. In conclusion, our proposed model hosting service Cosmic Collaboratory aims to revolutionize the SDA community by providing a specialized platform for sharing and accessing pretrained neural network models. This initiative will empower researchers to build upon existing work, respond rapidly to new challenges, and ultimately enhance the effectiveness of space domain awareness efforts. We hope to build a wide userbase community of AMOS attendees and usher in a new era of SDA collaboration.
Date of Conference: September 16-19, 2025
Track: Machine Learning for SDA Applications