280

Papers

10/10

Roadblocks Active

3

Connections

Negative triangularity (NT) and quasi-continuous exhaust (QCE) plasma shaping successfully eliminate large, damaging edge-localized modes (ELMs) across multiple European fusion devices — including JET operating with deuterium-tritium fuel. The mechanism is that these shapes suppress the pressure-driven instabilities at the plasma edge that cause ELMs, rerouting energy loss instead through gentler ballooning-mode turbulence. This matters because ELMs produce intense heat spikes on the divertor wall that current tungsten materials can barely survive; eliminating them is a prerequisite for continuous reactor operation.

██████████ 0.9 elm-control Preprint

ReadSaveConnections

This paper builds fast surrogate models — trained on detailed tritium diffusion and trapping simulations — for four plasma-facing components in Tokamak Energy's ST-E1 pilot plant design, then plugs those surrogates into a system-level fuel cycle model to rapidly evaluate how much tritium is lost, retained, or recovered across 80,000 seconds of pulsed operation. Tritium is the rare fuel for fusion and every atom that diffuses into a wall and isn't recovered represents both a fuel loss and a safety hazard; this multiscale approach makes it computationally feasible to optimize blanket and wall designs for maximum tritium recovery. Without tools like this, the question of whether a fusion plant can sustain its own fuel supply (tritium breeding ratio > 1) cannot be reliably answered during the design phase.

██████████ 0.9 tritium-breeding Preprint

ReadSaveConnections

Global gyrokinetic simulations reveal that turbulence does not stay where it is generated — it spreads radially into regions of the plasma that are linearly stable, and in doing so it drags zonal flows (self-organized plasma jets that normally suppress turbulence) outward with it. This coupling means that locally-measured turbulence suppression does not guarantee globally reduced heat loss, complicating predictions of overall confinement quality in reactor-scale plasmas. Understanding this mechanism is essential for building transport models that correctly predict how much heating power a reactor needs to reach ignition.

██████████ 0.9 turbulence-modeling Preprint

ReadSaveConnections

SOLPS-NN replaces the computationally expensive SOLPS-ITER plasma boundary simulation code — which typically takes hours per run — with a neural network trained on over 8,000 simulations, enabling millisecond-scale predictions of scrape-off layer conditions including whether the plasma is in a heat-reducing 'detached' state. Detachment is the main mechanism by which future reactors plan to keep divertor heat loads survivable (below ~10 MW/m²), so being able to predict and control it in real time is a critical engineering need. The surrogate achieves experimental agreement on detachment access and spans machine sizes from current devices up to DEMO, making it broadly applicable for design work.

██████████ 0.8 divertor-thermal Preprint

ReadSaveConnections

This theoretical review argues that localized structures in magnetized plasma turbulence — current sheets, magnetic flux ropes, reconnection sites — are the primary sites of energy dissipation and particle energization, not secondary phenomena embedded in a smooth turbulent background. For fusion, the relevance is indirect but real: these same structures appear in tokamak edge turbulence and can drive anomalous transport and fast-particle losses that degrade confinement. The paper provides a conceptual framework for why turbulence models that average over these structures may systematically underpredict energy losses.

██████████ 0.8 turbulence-modeling Preprint

ReadSaveConnections

This paper implements an implicit numerical scheme in the BSL6D code that couples fully kinetic ion physics with a simplified drift-kinetic electron model, removing the computational bottleneck caused by fast electron waves while still capturing ion-scale zonal flows that regulate turbulent transport. Current state-of-the-art kinetic simulations must either treat electrons very crudely or use prohibitively small time steps; this hybrid approach offers a middle path that retains the physically important ion dynamics at reduced cost. Validating such approaches in slab geometry, as done here, is a necessary first step before applying them to realistic tokamak geometry.

██████████ 0.8 turbulence-modeling Preprint

ReadSaveConnections

Two deep-learning architectures (a Koopman-based Transformer and a ConvLSTM-UNet) are trained to predict the time evolution of magnetized plasma instabilities from initial conditions alone, reproducing both the exponential growth phase and the nonlinear saturation of Kelvin-Helmholtz instabilities across a range of magnetic field strengths. The models preserve global conservation laws (energy, enstrophy) and capture Alfvén wave propagation — two criteria that previous purely data-driven surrogates often failed. Fast, physics-respecting surrogate models for MHD dynamics are a building block for real-time disruption prediction and avoidance systems in tokamaks.

██████████ 0.7 plasma-disruption Preprint

ReadSaveConnections

This reinforcement learning paper introduces Occupancy Reward Shaping (ORS), a method that uses a learned model of future state distributions to provide denser reward signals in settings where feedback is sparse — a common problem in long-horizon control tasks like managing a tokamak plasma. ORS is evaluated on three fusion control tasks and achieves a 2.2× performance improvement on average across 13 benchmark tasks without altering the theoretically optimal policy. Better offline RL methods for fusion control matter because data from disruptions and near-disruption events is rare and expensive to collect, making it hard to train control policies using standard online trial-and-error approaches.

██████████ 0.7 plasma-disruption Preprint

ReadSaveConnections

FlowRefiner uses flow matching — a generative modeling technique — to iteratively correct errors in autoregressive predictions of 3D turbulent flows, achieving state-of-the-art accuracy on forced isotropic turbulence and Taylor-Green vortex benchmarks. The key innovation is replacing stochastic denoising (which amplifies errors) with a deterministic ODE-based correction step, combined with a noise schedule that keeps refinement stable across multiple correction stages. Improved surrogate models for 3D turbulence have downstream value for fusion: faster, more accurate transport predictions reduce reliance on expensive full-physics simulations during reactor design optimization.

██████████ 0.7 turbulence-modeling Preprint

ReadSaveConnections

This theoretical paper proposes using muonic hydrogen — a proton orbited by a muon instead of an electron — to bombard boron-11, arguing that the muon's tight orbit screens the proton's charge and reduces the repulsive Coulomb barrier by several orders of magnitude at energies below 100 keV. Proton-boron fusion is attractive because it produces no neutrons (avoiding the activation problems of deuterium-tritium reactors), but its cross-section is extremely low at practical temperatures; if muon-catalyzed enhancement is as large as claimed, it could reopen aneutronic fusion as a viable path. The analysis is purely theoretical using standard WKB tunneling calculations, and the key open question is whether a muon catalyst can enable enough reactions before it decays to make the energy balance positive.

██████████ 0.6 q-engineering Preprint

ReadSaveConnections