Measuring for gas: computer vision will "see" impending storms and turbulence

Russian scientists have developed a new approach to analyzing images of gas and plasma flows using computer vision and neural networks. The method is suitable for studying such processes in the Earth's atmosphere and space, including the propagation of shock waves, the movement of aerosol particles, the formation of vortices and the evolution of cloud structures. When processing data from drones, airplanes, or satellites, algorithms can automatically identify such structures, determine their speed, direction of movement, size, and change over time, which makes the approach useful for analyzing weather events. For example, severe turbulence that is dangerous for aircraft, dust storms, or tracking the smallest space debris.

Analysis of physical phenomena at different scales

Scientists from the Faculty of Physics at Lomonosov Moscow State University presented an approach to analyzing images of gas and plasma flows using computer vision and neural networks. The technology allows you to automatically recognize and process arrays of large data obtained during high-speed panoramic shooting of streams. The development is designed to study complex processes on a wide range of scales. It can be used to recognize various dynamic structures visualized both in laboratory experiments and in the Earth's atmosphere.

As the scientists explained, the devices created on the basis of modern photonics make it possible to register images with high spatial resolution at a speed of millions of frames per second. The use of optical imaging methods makes it possible to study processes in the Earth's atmosphere, including the propagation of shock waves, the movement of aerosol, dust and smoke plumes, the formation of vortices, zones of turbulent mixing and the evolution of cloud structures.

"Combining high—speed optical panoramic visualization with computer vision algorithms and neural network models allows us to reach a fundamentally new level of experimental analysis and extract quantitative physical information from image arrays that were previously almost impossible to process manually," Igor Doroshchenko, senior researcher at the Department of Molecular Processes and Extreme States of Matter, told Izvestia.

When processing data from ground-based cameras, drones, airplanes, or satellites, neural network algorithms can automatically identify the structures being studied, determine their speed, direction of movement, size, and change over time. This makes the approach useful for analyzing weather events and as an additional data source for predicting dangerous processes, such as severe turbulence, dust storms, or rapid pollution transfer, the scientists said.

The technique will also find practical application in space: laboratory strip scans (a method for high-speed registration of fast—flowing processes) of moving particles can serve as a model for images of small-sized space debris, and recognition algorithms can be used to detect, classify, and evaluate trajectories. In addition, the approach is universal for a wide range of laboratory installations, including impact and wind tunnels, gas-dynamic and plasma stands, where fast automatic processing of large amounts of video data is required.

The research is carried out at the facilities of the MSU Faculty of Physics. To process arrays of digital video data, scientists use both classical image analysis algorithms and neural network models, including the popular YOLO and ResNet architectures.

The approach combining computer science, mechanics, photonics and optics makes it possible to automatically analyze the evolution of structures and fractures, determine flow regimes, track the dynamics of microparticles, and vortex formations in rarefied gases, liquids, and plasmas, said Irina Znamenskaya, professor at the Department of Molecular Processes and Extreme States of Matter at the Moscow State University Faculty of Physics.

Recipe for a planet: how life appeared on Earth

Scientists have suggested that the Earth could be much more autonomous.

Recognition in a wide range of tasks

The work of scientists is important primarily from the point of view of practical application, Alexander Khvatov, head of the Laboratory for Modeling Natural Systems at the ITMO Institute of AI, told Izvestia. According to him, processing hundreds of thousands of frames using computer vision and neural networks is not new in itself, but scaling it up has become a key element of the research, from microparticle analysis to tasks related to space processes.

"The results can be used for recognition and classification tasks in a wide range of areas of aerodynamics, heat transfer, and plasma modeling," the specialist noted.

There are already works in the world on the use of deep learning to study turbulence, plasma, fluid flows and high-speed survey data, said Yaroslav Seliverstov, a leading expert in the field of AI at University 2035.

"The novelty of MSU's work lies in the adaptation and integration of modern computer vision methods for the automatic analysis of large arrays of panoramic high—speed images of gas—dynamic and plasma processes, as well as in demonstrating the applicability of these methods to the tasks of monitoring microparticles and space debris," said the specialist.

Similar areas are being developed in the world's leading scientific centers, including NASA, ESA, a number of Chinese research institutes and other organizations, but the strength of the Russian development can be called the interdisciplinary combination of photonics, physics of high-speed processes and artificial intelligence within a single automated analysis system, said Yaroslav Seliverstov.

— In the global trend, this corresponds to the transition from simple visualization of physical experiments to their intellectual interpretation, when algorithms not only "see" the image, but also automatically extract physically significant information from it, — said the expert.

The work was published in the journal ActaAstronautica and was carried out within the framework of the state assignment of Moscow State University.​