Predicting how the atmosphere and the environment will change over time or how air will flow over an airplane can be very difficult for even the most powerful supercomputers to solve. Scientists rely on models to fill in the gaps between what they can emulate and what they need to estimate. But, as every meteorologist knows, the models often rely on partial or incorrect information, which can lead to bad predictions.
Now, researchers at the Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS) call the so-called “intelligent alloys” that combine the power of computational science with artificial intelligence to develop models that complete simulations to predict the evolution of science. The most complex systems.
In the paper published in Nature CommunicationsPetros Koumutsakos, Herbert S. Vinokur, Jr. Professor of Engineering and Applied Sciences and co-author Jane Bay, former Postdoctoral Fellow at the Institute of Applied Computational Sciences at SEAS, combines reinforcement practice with flowing counting methods. One of the most complex processes in engineering.
Reinforcement learning algorithms are a machine similar to BF Skinner’s behavioral conditioning experiments. Skinner, a professor of psychology at Harvard from 1959 to 1974, trained pigeons to play ping pong by rewarding an avian competitor who could pass the ball to his opponent. Gifts reinforce strategies such as cross-table shots, which often lead to point and delicious treats.
In intelligent combinations, pigeons are replaced by machine learning algorithms (or agents) that are learned by interacting with mathematical equations.
“We take an equation and play an agent’s learning game of completing parts of equations that we can not solve,” said Bay, who is now an assistant professor at the California Institute of Technology. “Agents add information from observations that the calculations can solve and then improve what they calculate.”
“In many complex systems, such as turbulent flows, we know the equations, but we can never have the computational power to solve them correctly for engineering and climate applications,” Koumutsakos said. “By using reinforcement practice, many agents can learn to complete state-of-the-art computational tools to accurately solve equations.”
Using this process, the researchers were able to estimate challenging turbulent currents that interact with solid walls, such as the turbine blade, rather than current methods.
“Every engineering system from offshore wind turbines to energy systems uses models for flow interaction with equipment and we can use this multi-agent reinforcement idea to develop, enhance and enhance models,” he said. Bay.
In the second paper published in Nature Machine Intelligence, Koumoutsakos and his colleagues have long used machine learning algorithms to speed up predictions in complex process simulations. Take morphogenesis, the process by which cells divide into tissues and organs. Understanding each stage of morphogenesis is crucial to understanding some diseases and organ defects, but no computer is big enough to capture and store each stage of morphogenesis for months.
“If a process happens in a matter of seconds and you want to understand how it works, you need a camera that takes pictures in milliseconds,” Koumutsakos said. “But if that process is part of a bigger process like morphogenesis that can take months or years and you’ve been trying to use a millisecond camera all that time scale – you have no resources.”
Koumutsakos and his team, which includes researchers from ETH Zurich and MIT, have demonstrated that AI can be used to create reduced representations of fine-scale simulations (equivalent to experimental images) by compressing information such as zipping large files. Algorithms can reverse the process and move the reduced image to its full state. Solving in reduced representation is faster and uses much less energy resources than performing calculations with full status.
“The big question is whether limited cases of reduced representation can be used to estimate full representations in the future,” Koumutsakos said.
The answer is yes.
“As algorithms learn the reduced representations we know, they do not need full representation to create reduced representations for what’s next in the process,” said Pantelis Vlachos, a graduate student and first author at SEAS. Paper.
Using these algorithms, researchers have proven that they can generate estimates thousands to millions of times faster than it takes to execute simulations with full resolution. As algorithms learn how to compress and compress information, they can create a complete representation of the estimate that can be compared to experiments. Researchers have demonstrated this approach to simulations of complex systems, including molecular processes and fluid mechanics.
“In one paper, we use AI to complete simulations by creating intelligent models. In another paper, we use AI to accelerate simulations through the magnitude of several orders. Next, we hope to explore how to combine the two. Are strong. There is a lot of room for innovation in the space between AI and computational science. ” Koumoutsakos said.
The Nature Machine Intelligence The paper was co-authored by Georgios Arampatzis (Harvard / ETH Zurich) and Caroline Uhler (MIT).