The Need for Speed: Revolutionizing CPS Design with AI

For more than two decades, the Need for Speed (NFS) video game franchise captured the hearts of young gamers across the globe with its high-octane thrills, heart-pounding car chases, and the adrenaline rush of illegal street racing. Yet for many, especially those who played the earliest iterations in the late ‘90s and early 2000s, NFS is most nostalgically memorable for its car customization features.

In the “Garage,” players could upgrade their cars piece by piece, swapping out everything from engine parts and tires to spoilers and neon lights to craft the hot rod of their dreams. And whenever a component was changed, a helpful dashboard displaying metrics like acceleration, top speed, and control would show the impact of that change on their car’s performance.

“Customization Evolution in NFS Games (1999-2022)”

In the real world, engineers only wish they had this kind of instant feedback. From cars to helicopters to submarines, vehicles today are nearly all cyber-physical systems (CPS), made as much of software as of steel. Traditionally, designing or modifying these complex systems requires extensive trial and error by expert engineers, resulting in significant expenditure of both time and resources.

But what if we could build a real-world version of the Need for Speed Garage? What if artificial intelligence (AI) and machine learning (ML) could lend a hand, making the design and testing process faster and more efficient while maintaining – or even improving – the quality and diversity of engineering solutions?

Well, to give away the punch-line: We did it.


In a recent project known as the Exploration Service for Synthesis and Evaluation of Novel CPS Emergent designs (ESSENCE for short), part of DARPA’s Symbiotic Design of Cyberphysical Systems (SDCPS) Program, Galois and its partners at Rutgers University, Purdue, and Charles River Associates developed a machine-learning-driven method and toolset for quickly iterating on novel designs of urban air mobility (UAM) systems, such as unmanned aerial vehicles (UAVs).

“Honestly, it functions pretty much exactly like the car modification in Need for Speed,” said Galois Research Engineer Sourya Dey. “I mean, UAVs don’t have spoilers, but they do have propellers and wings and motors for us to work with.”

Just like in the NFS Garage, users can modify their vehicle’s components in the ESSENCE system, harnessing the bespoke ML algorithm to receive immediate feedback on the impact of that change on the UAV’s performance. Alternately, users can ask the system to provide options for novel designs that hit certain performance metrics or are optimized for mission-specific tasks.

Want to see what will happen if you add that more powerful but heavier propeller? Pick that part from the vast component list, add it to your design, and see the immediate impact on the UAV’s maximum speed, altitude, stability, and fuel efficiency.

Our highest scoring designs derived primarily from quad-copter and “pickaxe”-like designs.

Or let’s say you need a UAV quad-copter design that can fly a particular path, carrying a cargo of a certain weight, at a particular altitude. Input your desired performance parameters and ESSENCE will generate numerous potential solutions that fit the bill, displaying clusters of designs with similar variables on a heat map.

In both cases, how each combination of components impacts the system as a whole is analyzed in moments by a machine learning algorithm that can make sense of huge amounts of data dramatically faster than the human brain.

“That kind of design would previously have to be done manually with other software,” Dey explained. “But we can get the ML algorithm to work alongside the human engineer as a codesigner, enabling us to search over a much wider space of design solutions, faster and at lower cost.”

Surrogate Modeling and Koopman Theory

Once promising designs are generated, they need to be tested – virtually, that is. 

“This isn’t easy,” said Dey. “Basically, testing every design is a really complex physics problem. You have this UAV design, and it has lift, drag, aerodynamics, all that stuff — so how do you know if it’s going to actually fly?”

This is where simulation via surrogate modeling and Koopman Theory comes in. Traditionally, simulating a UAV’s flight dynamics, lift, drag, and other factors is computationally expensive and slow. The surrogate modeling technique, particularly Galois’s novel use of Koopman Theory, used AI and deep learning to facilitate the creation of simplified models (surrogates) that could predict system behavior more quickly than full simulations. Think of it as a “virtual wind tunnel” that speeds up design validation. This not only made the tests faster but also enabled simulations under theoretical conditions that are difficult or impossible to replicate physically—such as the pressure of a submarine at unknown depths.

The result was DLKoopman: a complete piece of software (an open-source Python package) that can not only efficiently simulate the performance of novel UAV designs, but which is generalizable to a widespread variety of systems found in physics, engineering, and beyond, differentiating it from previous efforts to model complex systems and making the potential for practical applications nearly limitless.

The Larger Impact

The wider implications of this project can’t be overstated. By automating and improving the design and testing phases of cyber-physical systems, developers can create more diverse, efficient, and high-performing designs faster and cheaper. These innovations hold promise not just for UAVs, but for other industries requiring complex system designs, such as aerospace, automotive, and even medical devices. 

Seamless, rapid vehicle (and other cyber-physical system) customization is no longer confined to the realm of video games. With ESSENCE, Galois has developed an innovative solution to our “need for speed.”