Earlier this summer Intel® AI mentors partnered with researchers at NASA’s Frontier Development Lab (FDL) for an eight-week research sprint. During that time the researchers and their mentors used AI to tackle a host of issues, including prospecting on the moon, getting a better look at the sun, and predicting when and how floods will happen on Earth.
Moon for Good Challenge
It’s been over fifty years since the Apollo 11 astronauts landed on the moon. Humanity will return, and eventually we may return for good. The Moon for Good challenge envisioned how dwellers of a permanent lunar settlement would supply themselves with the necessary resources to maintain life on the moon’s surface.
Bringing resources from Earth to supply a permanent lunar settlement will be costly and difficult. To mitigate that, lunar settlers will have to source some of the minerals they need from the moon itself. Identifying sources of essential minerals (like metal meteorites) in the lunar surface could greatly contribute to the survivability of a potential human settlement.
There’s already a large amount of extant data about the lunar surface, and different maps can identify elevation, surface texture, temperature, or magnetic activity. The researchers and their mentors used six Intel® Xeon® processor-based general compute instances to preprocess approximately 35TB of this extant map data. This lunar surface data was not necessarily uniform or made to be combined. The teams had to struggle with different image resolutions and sizes, so pre-processing was one of the most challenging parts of their task. Eventually, Google Cloud enabled combining the datasets to fit into a single model, where each lunar variable can be visualized as a different topography in Google Moon.
By processing these data sets using Intel AI technology, the team was able to create a map that could be of use to potential lunar settlers. Not only did the map they created show metal deposits and resource locations, it also included other factors such as gravity, elevation, and temperature, all of which is vital information for a long-term human presence on the moon.
Solar Magnification Challenge
The sun isn’t just a static orb of light and heat. Indeed, Earth’s nearest star is dynamic, active, and has varying patterns, seasons, and activity that can affect space weather and conditions here. Studying long-term patterns in solar weather presents a challenge, though. In order to study long-term patterns, researchers need to use older data, sometimes dating back several decades. However, that older data is often low-resolution or of poor quality and doesn’t necessarily include the detail necessary to provide new insight.
The solution: Use today’s AI to improve yesterday’s data. The research team tackling FDL’s Super Resolution challenge used AI-powered super-resolution to get new perspectives on old images of the sun’s surface.
The researchers leveraged the power of the Google Cloud n1 standard-96 instance based on Intel Xeon Scalable processors for data pre-processing, and Google Cloud n1 standard-8instance with 32GB RAM based on an Intel Xeon processor for conventional model training. For inference, they used GCP’s N1standard-64 instance with 56GB RAM based on an Intel Xeon Scalable processor leveraging Intel’s Deep Neural Network Library (DNNL) and the PyTorch framework.
The result was a better picture of the sun over time. Older data and newer data combined to form a long-term picture of the weather, climate, and long-term trends on Earth’s nearest star.
Predicting and planning for floods has been a concern of human civilization for millennia and remain one of the most common and dangerous forms of natural disaster. The Floods Challenge team worked on answering questions like how much rainfall it takes for a given water system to reach its peak height. Also, given that information, how probable is it for a region to reach its peak height throughout the year?
The answers lay in the data. Specifically, in Google Cloud Deep Learning VM images which include Intel’s DNNL and Intel’s optimizations of TensorFlow and PyTorch. The team used GCP’s N1standard-96 instance with 960GB RAM for pre-processing, some training, and inference, and were able to build models of given geographical areas and predict when and under what conditions waters would peak.
The FDL’s Flood Challenge team is far from the only one using AI to predict disasters. A new European Space Agency (ESA) satellite launching this year is leveraging AI to predict floods in a given area. The satellite uses the Intel® Movidius™ Visual Processing Unit (VPU) which allows it to run vision-based processing models in real time, circumventing the time-consuming and often costly process of transmitting data to Earth and back. Because it doesn’t have to receive model outputs from the ground, it can run flood detection models powered by the Intel® Distribution of OpenVINO™ toolkit, allowing for much quicker flood detection and prediction.
Encountering a Larger World With AI
Examples like this are only a few instances of how AI, Intel technology, and partnerships with organizations like NASA FDL and Google Cloud are bringing about better knowledge of our world and what lies beyond it. AI allows us to mine past data for new insights, and to leverage data for transformative endeavors like future moon settlements or disaster prevention. Intel is proud to provide not only the hardware, but also the mentorship and leadership necessary to bring about new enthusiasm for AI and science. For more information on future work with NASA FDL, follow @IntelAI on Twitter.
For more details, check out the code and datasets of the projects.