A significant project combines state-of-the-art AI solutions with research on near-Earth space
The project also investigates processes that bring about the aurora borealis and simulates, with the help of artificial intelligence, plasma bursts in near-Earth space, utilising the computing power of a supercomputer.
Reconnection in Earth's magnetic field (Image: Vlasiator)
In late 2021, the Academy of Finland awarded a total of €700,000 in funding to the Distributed AI in Supercomputing and phYsics (DAISY) project, which combines the latest techniques in machine learning with the computing power of a supercomputer.
– Combining the processing of space data with artificial intelligence that utilises machine learning really brought the stars into alignment. We didn’t know how well they could be applied in tandem before starting to draw up the research funding application, says Professor of Computational Space Physics Minna Palmroth from the Department of Physics at the University of Helsinki.
Palmroth supervises the project with Professor of Computer Science Teemu Roos, who also heads AI-related education efforts at the Finnish Center for Artificial Intelligence (FCAI).
Supercomputers can be used to simulate phenomena occurring in near-Earth space and interpret the data gained from them. In use at the University of Helsinki is Vlasiator, a plasma physics model that simulates the conditions of this region of space, including ion-kinetic effects without numerical noise.
Using a supercomputer to simulate phenomena that occur in near-Earth space
– AI solutions that simultaneously achieve precision, reliability, scalability and explainability currently pose a major challenge in AI research, Roos says.
– To be at the forefront in the competitive field of global AI research, we need a sufficiently challenging application field where we can obtain practically unlimited quantities of extremely complex data.
Since the amount of data as such does not increase the difficulty, a sufficiently complex phenomenon to simulate was also needed. The space physics data to be processed does not originate in satellites. Instead, it is simulated. The final product of the AI analyses are novel representations of data, such as graphical model structures.
The simulations run on Vlasiator require enormous computing power. In the DAISY project, the simulations will be piloted and developed using the LUMI supercomputer, which was started up in Kajaani last year.
– At this scale, this kind of computing is not carried out anywhere else. We want to demonstrate that state-of-the-art AI innovations are not only produced by large corporations that collect data on their customers, Roos says.
Looking for magnetic reconnection
The first goal of the DAISY project is to determine the location of magnetic reconnection, where plasma is discharged on the nightside of Earth between two oppositely directed magnetic fields.
– Magnetic reconnection is a fundamental phenomenon associated with near-Earth space. It is the originator of both the aurora borealis and disturbances in electrical grids, Palmroth says.
Magnetic reconnection is difficult to detect, especially in three-dimensional modelling. Usually, it is seen, for example, as a spike in gauges connected to the grid. The data gained from the phenomenon is complex, but its location could be pinpointed with the help of artificial intelligence.
– Reconnection is a topological phenomenon, not just a pile of numbers, says Palmroth.
Artificial intelligence could be able to classify elements that belong together or that are separate in magnetic fields, and where the transition from one topology to another occurs, or where the reconnection precisely takes place.
– A deep-learning system learns to organise data into a presentable form, making it possible to observe when and where such topological changes occur, Roos says.
In addition to using computers with the capacity for classification, it must be investigated how the occurrence of the phenomenon can be described in a way that humans understand. This can be achieved with the help of new graphical models that can be used to produce topological and symbolic presentations.
– How do you model a phenomenon that is not just quantitative in nature? This is a fitting challenge for artificial intelligence, and we have good reason to believe that it can be solved, says Roos.
New ways of presenting data
After the technique has been finalised, uses can be sought for the graphical models it will produce from other fields of science as well.
– It would be a breakthrough if we were able to develop new models and thus be able to operate, as it were, in the middle ground between numerical and symbolic data, Roos says.
– The project entails a lot of even surprising components, and we can’t even know all of the potential applications to begin with, Palmroth notes.
– We have partnered with American neuroscientists and people from the New Children’s Hospital too, with whom we will try to uncover more broadly applicable methods from our findings. In neurosciences, they could be put to use, for example, in imaging the developing brain, Roos muses.
The two researchers, who work in neighbouring buildings on Kumpula Campus, are pleased about the project, which will provide both with groundbreaking research results and essential information.
– In inter-disciplinary projects, one party usually decides the topic of research, while the other one ‘spins the wheels’. In such cases, the results primarily benefit the party making the decisions. However, this project will provide us with the means to take big strides in both fields, Palmroth says.
The text is originally published on the University of Helsinki website.