NASA’s efforts to find life in the far reaches of the universe draw significant attention. Less known is a recent agency program to prevent asteroids from ending life here on Earth.
Three teams of researchers using GPU-powered deep learning spent the summer tackling asteroid-related challenges at NASA’s Frontier Development Lab, in partnership with the SETI Institute, a nonprofit organization devoted to researching life in the universe.
The FDL’s work was made possible by an “applied research accelerator” — a GPU-powered platform that helps researchers accomplish work that used to take six months or more in just six weeks.
The FDL itself is a response to the White House’s Asteroid Grand Challenge, an ongoing program that aims to get researchers to “find all asteroid threats to human populations and know what to do about them.” The emergence of GPU computing has enabled organizations like NASA and SETI to analyze large datasets that are critical to the space program.
FDL Director James Parr says the lab’s goal was to approach the challenge in two ways: apply machine learning techniques and technologies to planetary defense, and demonstrate the viability of the applied research accelerator to industrialize significant breakthroughs quickly.
A group of 12 standout graduate students was chosen for the internship, during which they were housed at NASA Ames Research Center in Silicon Valley, and worked on their respective projects at a nearby SETI facility.
Machine Learning to the Rescue
To do this, the FDL broke the group into three teams. Each took on a component of asteroid defense deemed suitable for a machine learning approach: assessing deflection technologies, modeling shapes from radar data and locating fallen meteors to determine their composition.
Their overall aim: answer three questions that will be critical to contending with a potentially hazardous asteroid:
What Is It Made of?
The students tackling this question designed an autonomous drone to find meteorites in the field. They used GPU-powered deep learning models to build an automated meteorite detection system, drawing from a database of 25,000 training images of meteorites along with a 15-million-image library. Parr said that while the 0.7-percent false positive rate that resulted is still too high, the path toward a meteor-finding drone is now clear.
What Shape Is It, and What Is Its Center of Mass?
Knowing the shape of an asteroid is critical to assessing potential deflection efforts. But generating a single asteroid’s shape has traditionally required long computer runs using 50,000 lines of legacy code, plus about four weeks of human-guided iterations, said Parr.
The FDL team applied GPUs and machine learning techniques to reduce the search for an asteroid’s spin axis to a few hours of computing. They then applied what Parr would only call “a cutting-edge method” to bypass legacy code until the final iterations. Parr called the initial results promising, with weeks of computing requirements reduced to milliseconds.
What Is the Best Way to Deflect It?
Previous efforts to analyze proposed asteroid deflection techniques only accommodated about four possible orbits. The FDL team used GPUs and machine learning techniques to build an analysis model that could be applied to 800,000 simulated orbits — an astounding improvement to preparedness efforts.
GPUs Fuel Space Program
Parr stressed that none of this progress would have been possible without the four TITAN X and eight Pascal architecture-based GPUs NVIDIA provided to the FDL. These enabled each team to tap large data sets using deep learning approaches.
“(GPUs) may be as significant as the microprocessor was to Apollo in the late ‘60s,” said Parr.
As much as GPU-powered advances may one day help to save us from celestial annihilation, a lot of work remains to be done.
Said Parr: “There are still many unanswered questions that have to be resolved before we can successfully stop an asteroid on collision with a highly populated region on Earth.”