How do you think AI and ML can aid fusion reactors? Any ideas?

These models of AI and ML can help researchers in the fusion process



Most physicists and researchers have been trying to create an efficient nuclear fusion reaction. But unfortunately, they have not succeeded yet. While coming to AI and ML, these technologies are helping many industries and applications do complicated things that humans are not capable of doing. So why not the neural nets and the GPUs add power to AI and ML to lend their hands in nuclear fusion too? 

Diogo Ferreira, who is a professor of information systems at the University of Lisbon’s Instituto Superior Tecnico in Portugal says that physicists can write equations, develop theoretical models, and can manipulate things mathematically. But AI has no limit, it can help with this, he adds. 

Diogo along with colleagues have collaborated and started working on a Joint European Torus (JET) in the UK. This is a study that lists down three diverse uses of AI, ML, and Deep Learning models that can be helpful in fusion research. He trained these models using diagnostic data collected from 48 sensors that are connected to the JET sector, called ‘bolometers’ which can collect data of power and radiation. Let’s know more about each model now. 

The one model predicts disruptions in a super-hot plasma. Depending on how the model is trained, it can either predict the likely nature of the disruption where the plasma escapes confinement, jolting equipment can reduce plasma’s temperature by ending the reaction. The other prediction is of the time at which the disruption can occur. 

Coming to the second model it detects the anomalies in the plasma. This model is trained only on reactions where disruptions did not take place, and can also reproduce the right experiments. The model can identify when and how the data diverges from that of a successful reaction. This can be very useful for the scientists and researchers to understand which kinds of disruptions are likely to occur less. 

The last one surrounds the visual representation of plasma radiation patterns. It can be used for performing brute-force, direct calculations, and take 20minutes for the reaction. This becomes another model for Diogo’s research that can produce similar images within seconds or even less at times. 

On the other side the researchers at the University of Washington have published a study describing a method that can use machine learning to predict the behavior of a plasma. This model uses a technique called regression. It gives scenarios that lead to nonsensical results, enabling it to use data, less computational power in less time. 

Chris Hansen, researcher of this model says that the model of the study used a single GPU to control a fusion experiment that has previously required several computers. With this model it can give out inputs quickly and can be helpful during the experiment. 

Similarly in another study, Stefano Markidis an associate professor along with his colleague Xavier Aguilar created a deep learning model that can solve one of the more computationally intensive steps of determining information on a plasma through calculating its electric field. This method showed to be faster as well as accurate too. 

But with advantages, comes the disadvantages too. AI and ML come with their set of cons in nuclear fusion systems But with the help of machine learning algorithms such as deep learning scientists can glean bits of what these models see and learn more about the physics of plasma and fusion.