How NASA Is Helping Humans Reach the Red Planet, Using GPUs

Estimated reading time: 4 minutes

It can take up to 10 months for a spacecraft to get from Earth to Mars. But the entire journey can be in vain if something goes wrong in the last six minutes.

That’s the amount of time the vehicle has to slow from its initial descent rate of 12,000 miles per hour to zero for a smooth landing. The process is incredibly complicated, and a lander only has one shot to get it right.

“Landing’s not something you can bail out on,” said Ashley Korzun, researcher in the Atmospheric Flight and Entry Systems Branch at NASA Langley Research Center. “Once you start, you’re going to the surface.”

A group of NASA scientists and engineers is working with colleagues from Old Dominion University and NVIDIA to simulate with unprecedented accuracy the physics needed to land the first manned Mars mission. To do so, they’re using the fastest supercomputer in the world, the NVIDIA GPU-powered Summit system at Oak Ridge National Laboratory.

“If we were to run the studies that we did on a conventional computing platform, one job would take nine months,” said Aaron Walden, research computer scientist at NASA Langley, in a talk at this week’s GTC DC — the Washington edition of NVIDIA’s GPU Technology Conference. “But on Summit, that job takes about a week to run, and we can do six jobs at the same time.”

Running simulations on Summit also allows the team to incorporate a much higher resolution than prior projects.

Sticking the Mars Landing

While NASA has sent humans to the moon and rovers to Mars, a human mission to the Red Planet is more than a decade away. Before a mission can launch in the 2030s, there are multiple challenges to solve: how to protect astronauts from space radiation, how to enable humans to live and work on Mars — and how to safely land heavy loads on the planet.

The planets only align every 18 to 24 months for travel between Earth and Mars, so astronauts will need to stay a while. To make a long visit possible, unmanned spacecraft will go first to drop off supplies.

Past Mars trips have used landers that weigh around a ton, allowing space agencies to rely on air resistance and a parachute to safely get a vehicle through the atmosphere and to the Martian surface.

“That paradigm completely breaks down when we go to the vehicle size needed to support human missions,” said Eric Nielsen, a senior research scientist at NASA Langley.

Yes, It’s Rocket Science

A manned vehicle will weigh tens of metric tons, at least. That’s 10x more than the Curiosity rover, which in 2012 became NASA’s heaviest lander to reach the planet. Mars’ atmosphere will absorb most of the vehicle’s kinetic energy as it hurtles down to the surface, but not nearly enough for a soft landing.

So NASA plans for the first time to use retropropulsion. Firing a lander’s engines in the opposite direction of the planet’s surface at supersonic conditions will create upward thrust that can slow down the heavy vehicle’s descent. Complex fluid dynamics will be at work, from tiny turbulent eddies just centimeters from the vehicle to larger airflow meters away.

Since the planet’s atmosphere varies so much from Earth’s, there are limited options to test the physics involved as a lander approaches Mars’ surface. That’s why NASA relies on high-resolution simulations to plan how to control the speed and orientation of the vehicle under different landing conditions.

Scaling Up Simulations on Summit

Nielsen and Walden, along with others at NASA Langley and NASA Ames, work with Korzun’s team to simulate the fluid dynamics involved in a Mars landing on the Summit supercomputer.

Using the computational power of 3,312 NVIDIA V100 Tensor Core GPUs, the team can run an ensemble of six simulations at once with NASA’s FUN3D computational fluid dynamics software.

“Working on Summit has been so enabling for us,” said Korzun. “We’re able to run problems that are large-scale, but still contain fine enough resolution to capture far more of the relevant physics than ever before.”

Each simulation produces around 200 terabytes of output data, Nielsen said. Hundreds of thousands of them will be run to test a range of possible conditions a vehicle could encounter during its descent to the Martian surface.

Using CPU-based supercomputers operating in a capacity environment, just one typical simulation would take months, Nielsen said. Moving to Summit has allowed the team to run several in a single workweek.

“It’s been pretty game-changing in terms of the learning cycle,” he said.

Simulation data informs vehicle design, but could also be used in an actual mission: A spacecraft’s landing software could compare its real-time diagnostics with the simulation database every half-second during its dive through the atmosphere to help fly the vehicle.