Originally published on October 25, 2023
This article is also available in German: Warum künstliche Intelligenz in der Fusionsforschung an Bedeutung gewinnt
Fusion reactions take place in plasmas with temperatures of many millions of degrees Celsius, i.e. in states of matter with charged particles (ions and electrons). On the way to a fusion power plant, it is of central importance to be able to generate certain plasma states in a targeted manner. Whether this can be achieved in an experiment can only ever be determined by indirect measurements – and that is anything but trivial. This is because a plasma is a highly complex, inhomogeneous and constantly changing entity that is also invisible to the human eye in the essential areas. That’s why experimental facilities like ASDEX Upgrade and Wendelstein 7-X are surrounded by an armada of high-tech measuring instruments that generate several gigabytes of measurement data during every second of a plasma discharge.
In order to be able to evaluate these enormous amounts of data, machine learning (ML) methods are increasingly being used. After all, that is precisely the strength of artificial intelligence (AI): If you feed it enough training data, it can recognise patterns in gigantic amounts of data and derive principles from them.
“We have built up quite a bit of expertise in this field at IPP over the past few years,” explains working group leader Dr. Udo von Toussaint. “And the number of researchers at the institute developing AI algorithms continues to grow.” The most recent example of the IPP’s successes: when the journal “Contributions to Plasma Physics” published a special issue on “Machine learning methods in plasma physics” in June with important current work on this topic, three of twelve articles came from the IPP. Dr von Toussaint was part of the editorial team for the issue.
Finding and correcting fundamental measurement errors
One of these papers deals with a fundamental problem of experimental physics that occurs particularly frequently in plasma physics: how to detect “outliers” in measurements – i.e. measured values that must be sorted out so that they do not falsify the result. In fusion experiments, for example, they often occur because the neutrons released in the process unintentionally hit diagnostic imaging instruments and cause results that have nothing to do with the physical phenomenon actually being studied.
Doctoral student Katharina Rath (IPP, Ludwig-Maximilians-Universität Munich) developed an algorithm for robust data analysis in a team with other researchers. The challenge: In simple experiments, outliers can usually be detected with the naked eye because they do not fit into the series of values; in plasma physics, however, multi-dimensional data clouds of measured values arise that defy simple assessment. Classical training strategies in machine learning also have problems dealing with outliers. “It is the strength of the IPP to develop strategies here that nevertheless work. We succeeded in this work with the help of so-called Student t-processes,” says Udo von Toussaint. What’s more, the algorithm should also be able to estimate missing measuring points in the future. In extreme cases, it could even be possible, for example in the event of a temporary failure of one of the many measuring devices, to interpolate its missing results from the values of the other diagnostics.
AI to control the plasma state
Another groundbreaking work in the field of machine learning, which has now been published by IPP and other researchers, deals with the live detection of the state of plasma equilibrium in stellarators such as Wendelstein 7-X. The aim is to achieve this ideal state of plasma equilibrium. In general, this ideal state of the plasma should be reached and maintained. “Because we cannot see the plasma as a whole, we have to rely here on derived measurands such as magnetic field strength, luminosity effects and temperature measurements on the wall of the vessel,” explains Udo von Toussaint. “Nevertheless, there is a very high probability that we are often off by a few centimetres when determining the shape of the plasma.”
Only if the position and shape of the plasma can be determined live during experiments it is possible to establish and maintain a suitable state of equilibrium by readjusting external parameters. For an adapted control, however, it is important to estimate the uncertainty of these data – otherwise the control could react too strongly to erroneous signals. Doctoral student Robert Köberl (IPP, TUM, TU Graz), together with other researchers, developed an algorithm for this purpose that can not only reconstruct the plasma state (the MHD equilibrium) but at the same time determine which error range results from the measured values. “We can’t yet calculate fast, robust and accurate at the same time, because that would clearly overtax the capabilities of today’s computers,” says Dr von Toussaint. “But the calculation of the error range can provide an important basis for decision-making in experiment control.” If the algorithm indicates a high deviation from the plasma equilibrium and a small error range, the experiment control would have to readjust. If, on the other hand, the error range is large, it would doubt the measured value and not react.
Physics and AI: stronger as a team
An increasingly important field is the consideration of existing boundary conditions when training and applying machine learning methods. In plasma physics in particular, there is extensive knowledge about the allowed states of plasmas, since they are subject to energy and momentum conservation, for example. IPP Director Prof. Eric Sonnendrücker (IPP) was recently awarded the Dawson Award for his fundamental work in this field.
However, it is often very difficult to consistently integrate these existing boundary conditions into the commonly used AI/ML methods. Here, IPP scientists Dr. Tomasz Tyranowski and Dr. Michael Kraus were able to significantly improve the long-term stability of plasma simulations for one of the fundamental model systems in plasma physics (Dr. Tyranowski has since followed a call to the University of Twente). It is about the Vlasov model, in which the so-called symplectic equation structure is now correctly taken into account in the AI algorithms.
“These examples show that machine learning methods are often perfectly suited to solving problems in plasma physics,” judges Udo von Toussaint. “But our findings can also help completely different scientific disciplines.” The fact that solutions from AI/ML research can often be used universally was recently demonstrated by the annual MaxEnt conference (International Workshop on Bayesian Inference and Maximum Entropy Methods in Science and Engineering), which took place at IPP in Garching in July 2023. This year, artificial intelligence and machine learning were a big topic.
MaxEnt brings together researchers who use probabilistic models – mathematical models based on random variables and their probability distributions – to solve problems from a wide range of disciplines: It’s about materials science and engineering, earthquake probabilities, medicine and, indeed, plasma physics. “We all learn from each other there,” says Udo von Toussaint. And so it happens that AI/ML algorithms from the IPP are also used in the geosciences and meteorology, for example – or vice versa.
Artificial Intelligence and Machine Learning
AI is usually understood as the ability of a computer to react independently to input or information from outside and to learn from it. Machine learning algorithms are an essential component of AI. These are essentially systems that are fed with training data and develop statistical models that enable them to recognise regularities and patterns. Machine learning encompasses many different methods. Its most prominent representatives are artificial neural networks and stochastic processes (such as Gaussian processes). The well-known AI tool ChatGPT is based on neural networks.