Sage pioneered and demonstrated cyberinfrastructure to support scientific user code running securely on remote systems "at the edge."  This enabled students and scientists to process full resolution instrument and sensor data, using AI to create new scientific measurements. One user might use camera input and an AI model to count pedestrians; another might use the inputs to classify weather patterns. This flexible, programmable approach to processing data, beginning directly at the sensor, represents a computing continuum run-time system linking cloud computing to the edge.

The Sage open source software framework integrated and operated a national fleet of over 120 AI-enabled nodes. The resulting body of software, community code development processes, and scientific experiments strengthened and deepened the Sage software reliability, ease-of-use, and critical system functions including cybersecurity, reliable data movement, and edge application management.

Sage also created open hardware designs for AI computation at the edge, including both outdoor-hardened and commodity rack-mounted nodes deployed across the country, providing the scientific community with a robust reference platform for supporting AI computation linked to their selection of instruments and sensors.   

Partnering with ARM Research, Sage built the edge programming environment using robust, industry standard software layers including software containers (Docker), process orchestration (Kubernetes), AI and computer vision toolkits (TensorFlow, PyTorch, and OpenCV), Python notebooks (Jupyter), and reliable messaging (RabbitMQ). Integrating these cloud-native software tools into the edge AI environment provides users with a familiar AI programming environment. Security features built into the Linux operating system enabled novel, advanced Sage cybersecurity policies to ensure, for example, that user codes could introduce network vulnerabilities, relying exclusively on Sage encrypted data messaging layers.  All of the code developed in the partnership is open source and published on GitHub.

Sage also created an open Edge Code Repository (ECR) to build and store reusable codes to be deployed to the edge, empowering students to develop new or improved features atop tested applications, and scientists to share AI codes and techniques. Each code in the ECR is automatically built and tested in the cloud, with full documentation. Concurrently, the Sage portal provides tools for integrating new sensors, viewing live data streams, downloading datasets, and scheduling AI@Edge jobs---deepening the catalog of integrated services for Sage users.

Early edge-AI adopter communities–essential to validating the Sage cyberinfrastructure, architecture, software interfaces, and usability–represent diverse domains including computer science, climate and atmospheric science, engineering, biology, microbiology, ecology, and sociology---spanning five NSF directorates. Each provided unique design input and usability feedback as Sage was built and deployed, contributing to a unique national AI cyberinfrastructure.  Disciplinary partners include the U.S. National Ecological Observation Network (NEON) that samples the biosphere at a continental scale; the urban science community launched by the Array of Things to measure urban environments, air quality, and human activity; the Atmospheric Radiation Measurement (ARM) User Facility measuring key climate processes; the High Performance Wireless Research and Education Network (HPWREN) that provides real-time measurements of natural phenomena in southern California, from wildfires to earthquakes; the Oregon Hazards Laboratory (OHAZ) that participates in the ShakeAlert and ALERTWildfire network; the Nature Conservancy that measures wildlife; and the Great Lakes Indian Fish and Wildlife Commission (GLIFWC) to study wild rice in partnership with Ojibwe tribes. These partners leveraged Sage and AI@Edge to provide new, advanced data products for research not previously possible.

Today, Sage is a robust, distributed national network of nodes deployed in both harsh outdoor settings and in protected instrument huts.  Within each partner’s privacy policy and legal constraints, Sage collects image and audio training data, creating training libraries  for collaborators developing AI models.  These libraries have been used to train AI models for such tasks as detecting floods, estimating snow coverage, and estimating biodiversity with audio. More than 1 billion data records and over 25 terabytes of training data have been collected. Online tutorials, sample codes, and a dynamic online community (via discussion forums and Slack) have enabled users to download and install Sage without assistance from the core development team.  Users have added new instruments to Sage, such as lidar, software-defined radios, an amateur radio telescope, an ultrasonic anemometer, and camera pan-tilt units to explore AI for cyberphysical control systems.

Sage will continue to impact science.  Computer scientists can explore novel AI approaches and approaches for evaluating AI model performance, while cientists in other disciplines can focus on developing and using machine learning algorithms to analyze high-volume sensor data streams, relying on Sage critical cyberinfrastructure for data movement, networking, cybersecurity, and compute resources.  Dozens of students have contributed to Sage, gaining research skills at the nexus of cyberinfrastructure and domain science applications, inspiring the pursuit of STEM careers by providing a platform for exploring measurement-based science questions related to the natural and built environments, creating applications that focus on social or humanitarian goals relevant to them.

Last Modified: 05/29/2024
Modified by: Peter H Beckman