DeepMind taught AI to control nuclear fusion

DeepMind taught AI to control nuclear fusion

After revolutionizing chess, go game, programming or biology, these real AI wizards are now dealing with nuclear fusion.

One of the major obstacles to further progress in the question of nuclear fusion is the control of the reaction. Because if the basic scientific theory is fully mastered today, it is still far from practical; We heard about significant progress China And inside England, But a lot more work needs to be done before we can believe we will absorb this reaction. But the latest work situation can change with a very different approach: ask an artificial intelligence to do it for us!

This idea comes to us directly from the wizards at DeepMind who specialize in AI, founded by the brilliant Demis Hasabis. In recent years, this cousin of Google (DeepMind, based on Google’s parent company Alphabet) has set up several labs, all of which have one thing in common: they seek to solve the problem of basic research on artificial intelligence. .

The company started by creating a name for itself in games with such projects Albostar, Alpha Chess Where Albaco Today it has reached an undeniable level of inhumanity. Gradually, DeepMind began to explore AI’s interest in revolutionary fields that were certainly less fun for the general public, but at the scientific level.

Its teams have produced some of the most impressive and equally solid works in recent years. Within a few years, its researchers, for example, overcame the general theory Looking for solutions, Began work that could redefine computer science and programming in general, and proposed a system that would revolutionize structural biology altogether – that is. But they were not done, so far. From now on, these real algorithm wizards will keep their eyes on the long-awaited nuclear fusion of our civilization.

JET tokamak is another very useful testing platform for verifying selected development axes for ITER. © Christopher Roux (CEA-IRFM) / EUROfusion

Plasma, a storage but capricious partner

In a nutshell, it is more or less the same Cook in a little sun In the most expensive pressure cooker in history. The technical challenge is enormous; After all, it involves controlling a heated plasma to a demonic temperature, beyond the index threshold of 100 million degrees Celsius. The problem is that science today does not have the capacity to withstand such furnaces.

To deal with this size problem, ITER Other tokamakes are based on a brilliant idea developed in the 1950s. If no object can withstand this temperature, just … do not use any physical object! Or more precisely, to make sure that he never, under any pretext, comes into contact with plasma.

To achieve this, researchers are using massive electromagnets to create a very powerful magnetic field, which is capable of controlling plasma at good distances from the walls of the reactor. Simple in theory, but much easier than possible. In fact, the parameters of this magnetic field must be controlled very precisely to improve the blockage of the plasma, and hence the efficiency of the furnace.

Simulate for better experimentation

This is a barrier that significantly reduces the growth of Tocamax. Every time researchers want to improve the coil configuration they face a more serious engineering challenge. You need to properly model this new confirmation and then reconnect several tons of electromagnets and all other related components. Usually a long process, and without the slightest guarantee of results.

This is where DeepMind comes in, which uses AI to solve this dilemma. In fact, it is a scenario most suited to the capabilities of AI; It can handle many excellent parameters to achieve results very quickly that require years of hard work by researchers. So Deep Mind has developed a system capable of manipulating the plasma inside the chamber in real time, first in simulation, and then in a real reactor, modifying the function of the coils. This flexibility makes it possible to improve the stability and yield of the reaction.

© DeepMind

A Schizophrenic AI in the Scientific Service

To achieve this, researchers have to measure almost a hundred different parameters ten thousand times per second. The hell rate, like the temperature of the reactor, is much higher than all the systems made by the company until then. This forced DeepMind to change its approach. What makes this system special is that it is not one, but is made up of two sub-networks of virtual neurons.

These interact, however, and operate independently. So we have a kind of schizophrenic AI, a network on one page “Important“Relatively stable learns through repetitions, another says second network”Actor“It simply came to our notice then.

The system was trained at the Swiss Plasma Center’s TCV for Tokama. This is a somewhat specialized machine as it has a variable configuration that allows researchers to adjust room parameters between two experiments. They hope to identify the optimal configuration to be reused within the framework of ITER and allies while providing evidence of a long-awaited concept.


Other rapid and concrete progress is expected

At the end of this tutorial on different and different configurations, DeepMind’s AI was able to actually predict the magnetic field manipulations required to achieve different plasma configurations. The good news is, these results can be verified in practice; During real-life experiments on TCV, researchers were able to obtain plasma configurations that were very close to simulations.

This does not mean that we will have access to business mergers in the future. On the other hand, it’s a worthwhile decision. “Remarkable“And”Very positive”By experts. This work has proven to be a reliable and highly efficient partner in exploring DeepMind’s new AI reactor dynamics, which will significantly accelerate the research and development of other sources of concept. This is a technology that typically enhances the performance of engines such as the JET, another test to provide ITER service to Tokamak, which recently broke the power generation record.

Admittedly, it is important to ensure that these results, specific to TCV, apply to other tocopherols as well. But this system is undoubtedly the most promising research tool that can be democratized in the coming years. There is reason to be excited as the first plasma of ITER, planned for 2025, approaches.

Research article available Here.

Leave a Reply

Your email address will not be published. Required fields are marked *