Research combines artificial intelligence and computer science for accurate and efficient simulations of complex systems

Predicting how the climate and environment will change over time or how the air moves over an airplane is too complex for even the most powerful supercomputers to solve. Scientists rely on models to bridge the gap between what they can simulate and what they need to predict. But, as all meteorologists know, models are often based on partial or even erroneous information, which can lead to poor predictions.

Now, researchers at Harvard’s John A. Paulson School of Engineering and Applied Sciences (SEAS) are forming what they call “smart alloys,” combining the power of computer science with artificial intelligence to develop models that complement simulations to predict the evolution of science. the most complex systems.

In an article published in Nature CommunicationPetros Koumoutsakos, Herbert S. Winokur, Jr. Professor of Engineering and Applied Science and co-author Jane Bae, former postdoctoral fellow at the Institute of Applied Computational Science at SEAS, combined reinforcement learning with numerical methods to calculate turbulent flows, one of the most complex processes in engineering.

Reinforcement learning algorithms are the machine equivalent of BF Skinner’s behavioral conditioning experiments. Skinner, the Edgar Pierce professor of psychology at Harvard from 1959 to 1974, trained famous pigeons to play ping-pong by rewarding the avian competitor who could peck a ball past their opponent. Rewards reinforced strategies like crossfire that often resulted in a point and a tasty treat.

In smart alloys, pigeons are replaced by machine learning algorithms (or agents) that learn by interacting with mathematical equations.

“We take an equation and play a game where the agent learns to complete the parts of the equations that we can’t solve,” said Bae, who is now an assistant professor at the California Institute of Technology. “Agents add information from observations that the calculations can solve, and then they improve on what the calculation did.”

“In many complex systems like turbulent flows, we know the equations, but we’ll never have the computing power to solve them with enough precision for engineering and climate applications,” Koumoutsakos said. “Using reinforcement learning, many agents can learn to complement advanced computing tools to solve equations accurately.”

Using this process, the researchers were able to predict difficult turbulent flows interacting with solid walls, such as a turbine blade, more accurately than current methods.

“There is a huge range of applications because every engineering system, from offshore wind turbines to energy systems, uses models for the interaction of the flow with the device and we can use this idea of ​​multi-agent reinforcement to develop, increase and improve models,” Bébé said.

In a second article, published in Intelligence of natural machines, Koumoutsakos and his colleagues have used machine learning algorithms to speed up predictions in simulations of complex processes that take place over long periods of time. Take morphogenesis, the process of cell differentiation into tissues and organs. Understanding each stage of morphogenesis is essential to understanding certain diseases and organ defects, but no computer is large enough to image and store each stage of morphogenesis for months.

“If a process takes place in seconds and you want to understand how it works, you need a camera that takes pictures in milliseconds,” Koumoutsakos said. “But if that process is part of a larger process that takes place over months or years, like morphogenesis, and you’re trying to use a millisecond camera over that whole time scale, forget it – you’re missing out. of resources.”

Koumoutsakos and his team, which included researchers from ETH Zurich and MIT, demonstrated that AI could be used to generate reduced representations of small-scale simulations (the equivalent of experimental images), by compressing the information almost like compressing large files. Algorithms can then reverse the process, returning the reduced image to its complete state. Solving in the reduced representation is faster and uses far less energy resources than performing calculations with the complete state.

“The big question was whether we could use limited instances of reduced representations to predict full representations in the future,” Koumoutsakos said.

The answer was yes.

“Because the algorithms have learned from the reduced representations that we know are true, they don’t need the full representation to generate a reduced representation for what comes next in the process,” said Pantelis Vlachas, a student graduate at SEAS and first author of the paper.

Using these algorithms, the researchers demonstrated that they could generate predictions thousands to a million times faster than it would take to run the simulations at full resolution. Because the algorithms have learned to compress and decompress information, they can then generate a complete representation of the prediction, which can then be compared to experiments. The researchers demonstrated this approach on simulations of complex systems, including molecular processes and fluid mechanics.

“In one paper, we use AI to complement simulations by building intelligent models. In the other paper, we use AI to speed up simulations by orders of magnitude. Next, we hope to explore how to combine these two “We call these methods Smart Alloys because the fusion can be stronger than each of the parts. There’s a lot of room for innovation in the space between AI and computational science.” Koumoutsakos said.

the Intelligence of natural machines the article was co-authored by Georgios Arampatzis (Harvard/ETH Zurich) and Caroline Uhler (MIT).

Comments are closed.