The towering importance of data
The most important product of ITER is data, which will be used to produce the information needed to build models for DEMO and commercial reactors—and much more.
But ITER is also a consumer of data. Raw data and extrapolations from previous fusion experiments help scientists and engineers develop simulators that play a crucial role in design phases, and raw and processed data feed into analytical tools that will serve the equally important role of interpreting events within the plasma during operation.
Artificial intelligence is providing a new path to speed and efficiency in the development of models and analytical tools. Given the right training data, artificial neural networks can learn to spot patterns of input data that produce a given output.
The Science, Controls & Operation Department is ultimately responsible for all scientific data at ITER. Within that department, Simon Pinches leads the Plasma Modelling & Analysis Section—a team that develops tools to maximize the value of information.
Collecting input to simulate critical systems
"As an example, we have started to simulate diagnostics," says Pinches. "These simulations will grow into synthetic diagnostics that can be used to predict what will be seen on the real sensors, which will in turn help with the design of the diagnostic systems and in building the analytic tools that will interpret measurements. Because these simulators are so important to the project, we aim to have good models for all the diagnostics about five years before they're needed."
To make sure they do not miss important features, the Plasma Modeling & Analysis Section has been soliciting input. Within the ITER Tokamak Physics Activity, for example, is a Diagnostics Topical Group with a specific subgroup on synthetic diagnostics. Scientists from within the ITER Members have joined and have already provided useful ideas.
One area of special interest is understanding what sensors will detect at the approach of high-confinement mode (H-mode) in the plasma. "We have a Monaco Postdoctoral Fellow, Anna Medvedeva, who is looking into this," says Pinches. "We want to achieve high-confinement mode as soon as possible, since the plasma retains its heat much better once it has reached this state. To help us know when we are close to high-confinement mode, Medvedeva is modelling what we can expect the sensors to tell us."
In addition to simulating diagnostics, the Section is developing simulators of the different external heating systems because they strongly influence what goes on in the plasma. But there is one potential showstopper: whether they serve as a foundation for diagnostic or heating systems simulators, very detailed models suffer from the same downside—they take time to develop and can take an excruciatingly long time to run.
The challenge is not unique to nuclear fusion; meteorologists face the same problem. The sheer number of variables in traditional weather models and the complexity of their interactions are too great for even the latest processing technology to overcome. Fortunately, artificial intelligence is providing a new path to speed and efficiency—given the right training data, artificial neural networks can learn to spot patterns of input data that produce a given output.
The Plasma Modelling & Analysis Section intends to use this approach to transform input into output several orders of magnitude faster. "Our Monte Carlo models take quite a lot of time to run," says Pinches. "So we're examining all the different input data for all the different cases we think some of the systems will experience in the lifetime of ITER and looking at the data that comes out. We'll use this to train a neural network equivalent that can do the same thing incredibly quickly. This will allow us to run our simulations without having to wait for these complex calculations."
Producing the right data to train analytical tools
The team will apply the same thinking to the analytical tools. Using models that take into account not only direct measurements, but also the uncertainties of the measurements, analytic software will determine the state of the plasma. Examples of uncertainties include the angle of a mirror being slightly off, or a minute error in the position of a sensor.
"We will need a lot of data to create the neural network in the beginning," says Pinches. "This data will essentially 'teach' the systems to be able to assess the state of the plasma during operations. We plan to generate synthetic data, artificially produced, to train our network to behave like our complicated processing tools. We have to be careful to select the right data to train our neural networks. If we want to be able to spot a certain event in the plasma, we have to show it lots of examples of 'this is what it looks like, now can you go off and find it?'"
Inspiring the future through data and insight
These are just a few examples that illustrate the colossal importance of data in getting ITER off the ground. But while the project relies heavily on output from previous experiments, on balance it will produce more bits and bytes than it takes in—and that is where ITER is expected to have its most lasting impact.
return to the latest published articles