By combining a deep understanding of plasma physics with machine learning techniques, DIFFER researchers developed a new ultrafast neural network model of the turbulent plasma in a fusion reactor. The neural network can accurately predict heat and particle transport in the fusion reactor up to 100.000 times faster than before: a vital tool to optimize the performance of future fusion power plants.
Fusion reactors are fuelled by a plasma: a hot, ionized gas of hydrogen isotopes that fuse together at extreme temperatures to form helium and release clean energy. The behavior of the plasma is not easy to predict: the charged plasma particles respond not only to the magnetic field that keeps them trapped inside the reactor, but also to the electromagnetic fields they create themselves through their own motion. That makes predicting a fusion plasma in order to optimize its state a difficult but rewarding problem to tackle.
From days to tens of seconds
Current nonlinear models to describe the complex turbulent behavior of the plasma are very accurate, but require a lot of supercomputing time. Depending on how detailed the model is, this can range from 10.000 to 10 million CPU hours to simulate 1 millisecond of plasma behavior. Models with reduced physics complexity – but still valid for wide regimes in nuclear fusion plasmas – can predict tokamak plasma behavior within a few days using standard computers. By using neural networks, DIFFER-researchers Karel van de Plassche and Jonathan Citrin managed to reduce the simulation time even further, down to only a few tens of seconds.
‘Our model predicts turbulent transport with the same accuracy as the reduced physics models, but 100.000 times faster,’ Van de Plassche says. The researchers used supercomputers to calculate a database of almost a billion results of the reduced physics model, corresponding to the plasma state that could arise under different circumstances. ‘Subsequently, we used this database to train a neural network how the different inputs in the model, like electron and ion temperature gradients, are linked to resulting flows of plasma particles and heat. These relations are then fed into an integrated model which predicts the tokamak temperature and density evolution.’
Scenario design and control
For a series of studied test cases, the simulations made with the new model accurately predicted the turbulent transport of energy and particles in the Joint European Torus (JET) tokamak within seconds. ‘This considerable time-saving paves the way for extensive scenario design for fusion experiments, and speeds up the interpretation of the results,’ Citrin explains the relevance of the research. ‘Especially for ITER, accelerating the experimental campaigns towards the project goals through fast and accurate simulations will be vital. To achieve this type of full-device modelling, our neural network approach can be extended to other tokamak physics components beyond turbulent transport.’
‘For fusion reactor experiments, perhaps the most important boundary condition is to make sure you don’t break the machine,’ Van de Plassche says. ‘With fast and accurate simulations, it becomes possible to explore a larger operating space. You can even try the more risky settings to find out what would happen if you go out of bounds, which is something you would never do with a real reactor.’ ‘And this technology is a prerequisite for commercial reactors,’ Citrin emphasizes. ‘Since in such reactors, due to physical constraints the diagnostics will be more limited than in an experimental setting, fast, high quality simulations will be crucial for their control systems.’
Open Access publication
Fast modeling of turbulent transport in fusion plasmas using neural networks, Physics of Plasmas 27 (2020)
The neural network research in this paper is a good example of open science. Not only is the publication available as a pre-print, the underlying dataset, the machine learning code and the computational model are also freely available.