The need for clean and cheap energy sources is a stringent problem given the hazards of climate change and the socio-economic global context. One potential solution is controlled thermonuclear fusion, in particular fusion achieved in tokamak devices. The first tokamak was built almost 70 years ago, and, from that moment on, there was a slow but steady pace of progress toward the achievement of a hot, stable plasma that is confined and capable of producing more output than input energy via the fusion process. Unfortunately, we are still far from reaching that goal. While there are multiple reasons for this slow advancement, turbulent transport is, arguably, one of the main obstacles. This is no coincidence since turbulence, in general, is regarded as one of the greatest unsolved problems in all of science.
The difficulties can be understood from the physical picture of transport. A tokamak device is donut-shaped and contains a hydrogen plasma that is confined by a strong, mostly toroidal, magnetic field. Charged particles are expected to gyrate across and move along the magnetic field lines. A small, diffusive, and relatively harmless “neo-classical” transport is possible from the synergy between collisions and particle trajectories. Small fluctuations in the plasma draw energy from the density and temperature gradients and, through an instability mechanism, develop strong, chaotic, drift-type electromagnetic fields. The latter represents plasma turbulence. The electric and magnetic fields couple and cause the E×B drift on all charged particles. The turbulent nature of this motion leads to radial pinches and diffusions that are orders of magnitude larger than the neoclassical ones. The consequence is the degradation of confinement: particles and heat are lost from the core of the reactor, potentially leading to wall damage or even disruptions.
It is clear today that, to achieve and maintain controlled thermonuclear fusion in future tokamak devices (ITER, DEMO), we must mitigate such scenarios, and for that, a sound understanding of turbulence evolution and associated transport must be attained. In technical terms, we need theoretical and computational tools that are able to predict turbulence and its impact on plasma dynamics quickly and accurately.
The state-of-the-art approach used nowadays for this problem is gyrokinetic theory (GK). It is a kinetic framework reformulated in the gyro-center approximation that can describe a magnetized plasma at small scales, capturing turbulent phenomena. This approach is implemented in sophisticated numerical codes that solve the inherent system of partial differential equations. While GK simulations have obtained reasonable levels of agreement with experimental data, they have a serious drawback: the numerical effort required is large. Thus, the characterization of the correlation between plasma parameters and transport using GK is prohibitive for many computer architectures and the method is unsuitable for fast predictions, even more so for real-time applications.
On another scientific front, we have witnessed in the past few years the rise of artificial intelligence (AI), in particular machine learning (ML) techniques. ML promises that, if enough data is available, it can learn from that data and make inferences about various mathematical problems. Recently, machine learning tools have found their way into the domain of fusion plasmas due to their most appealing feature: the computational speed.