A few months ago, the largest sunspot appeared in the sun, which we have seen in the last 24 years. This monstrous spot was visible to the naked eye (that is, without an approximation, but with protective glasses) and generated more than a hundred flashes. The number of spots on the sun changes cyclically every 11 years, increasing and decreasing. Right now, the sun is in the most active part of this cycle: we expect a lot of spots and many flashes in the coming months.
People, as a rule, are frightened by the destructive power of solar flares - there is a chance that one day a powerful explosion at the Sun will hurl a ton of energy particles in our direction and burn our communications satellites. But no one thinks that we can predict such flashes, like the weather, and therefore prevent possible damage. But how to predict a solar flare?
One way is to use machine-learning programs such as artificial intelligence, which automatically extracts data from experience. These algorithms are constantly improving their mathematical models when new data appears. But in order to learn well, algorithms require large amounts of data. Scientists did not have enough data about the Sun until 2010, until the Solar Dynamics Observatory (SDO) was launched, the sun-observing satellite, which sends about one and a half terabytes of data to the earth every day — more than this satellite, did not send data no device in the history of NASA.
Solar flares, as is known, is a complex active process. They occur in the solar atmosphere above the sunspots located on the surface of the sun. Sunspots, which usually come in pairs, act as bar magnets - one spot as the north pole, the other as the south. Considering that there are a lot of sunspots on the Sun, different layers of the Sun rotate at different speeds, the Sun itself has a north and south pole, the magnetic field becomes extremely saturated. As a result, magnetic fields appear, twisted like an elastic band, which release a lot of energy in the course of their existence. So there is a solar flare. Sometimes twisted fields do not flash, sometimes flashes appear from innocuous-looking spots, sometimes giant sunspots produce nothing.
We do not know how solar flares arise from the point of view of physics. We have - we know that flares are magnetic - but we do not know how they release so much energy at such a speed. In the absence of a final physical theory, the best hope for predicting solar flares lies in processing our gigantic data sets in search of clues.
There are two main ways to predict solar flares: numerical models and statistical models. In the first case, we take the physics we know as the basis, compose the equations, run them in time, and get the prediction. In the second, we use statistics. We answer the questions: what is the probability that an outbreak will appear in an active region with a giant sunspot? And what is the likelihood that this will happen with a small spot? As a result, there are giant data sets, full of details: the size of sunspots, the strength of the magnetic field. Scientists then look for connections between these details and solar flares.
Machine learning algorithms can put an end to this. We use machine learning algorithms everywhere. Biometric clocks use them to wake us up when needed by our body. They are better than doctors predict rare genetic diseases. They define the pictures that inspired scientists throughout history. Scientists consider machine learning algorithms to be universally useful because they can understand non-linear data, and for large arrays it is almost impossible for people to do. But a lot of models are non-linear, so these algorithms are increasingly being used in all areas.
Scientists use machine learning algorithms to predict solar flares based on a giant SDO data set. For this, we had to build a database of all active regions that SDO had ever seen. Since this is historical data, we already know whether these active regions have flared up or not. The learning algorithm analyzes the details of the active regions — the size of sunspots, the strength of the associated magnetic fields, and their torsion — to reveal the general characteristics of the actively flashing region.
To do this, the algorithm begins with an assumption. Suppose, at first, he assumes that a tiny sunspot with a weak magnetic field will produce a giant flash. Then checks the answer. Oops, no. Then the algorithm rebuilds its assumption. Next time he will enter in a different way. With trial and error, with hundreds of thousands of assumptions and tests, the algorithm gradually improves the accuracy of its predictions. It can be applied to real-time data, and he continues to learn.
Continuing to work in this direction will provide us with a better understanding of the upcoming solar flares. So far, scientists have shown that machine learning algorithms are better or, at worst, the same as statistical or numerical methods. And that's cool, actually. Such algorithms, which can work without the participation of people, simply by looking at huge data arrays, will be infinitely useful - and the farther, the more - in various fields. The most curious thing is that the same algorithms that make predictions of solar flares can work with genetic diseases and their definition.
And what if there is more data? Who knows. Although we already have a lot of data about the Sun - SDO has been working for four and a half years - there haven't been so many solar flares since then. Because we are in the quietest solar cycle of the century. There is a reason to continue collecting data.
The article is based on materials .
Comments
Post a Comment