292

Papers

10/10

Roadblocks Active

6

Connections

Injecting trace tungsten into DIII-D plasmas lowered electron temperature enough to stabilize a class of turbulence called trapped-electron modes, which in turn cut ion heat losses roughly in half and doubled toroidal plasma rotation. This matters because the same impurity seeding used to protect the divertor from heat overload simultaneously improved core plasma confinement — a rare case where an exhaust solution also helps the core rather than hurting it. The result is based on a small set of six discharges, so broader confirmation is needed, but the mechanism is physically clear and testable.

██████████ 0.9 plasma-wall Preprint

ReadSaveConnections

A neural network trained on microwave diagnostic data from the plasma edge can predict the first ELM crash 100 milliseconds before it occurs — enough lead time to trigger mitigation hardware. The model treats ELM prediction as a survival analysis problem, outputting a probability distribution over when a crash will happen within the next 150 ms using only 50 ms of input data. This is a proof-of-concept on DIII-D data with incomplete reporting of dataset size, so real-time deployment would require further validation, but the prediction window is practically meaningful.

██████████ 0.9 elm-control Preprint

ReadSaveConnections

Experiments at ASDEX Upgrade mapped out the reproducible sequence of events that unfolds when a shattered pellet is injected to stop a plasma disruption: fragment arrival, plasma movement, radiation collapse, thermal quench, and current decay follow a consistent timeline whose shape — convex versus concave current trace — reliably indicates whether mitigation was effective. Fragment size and velocity from three different shatter head geometries were varied systematically to isolate their effects on this sequence. This level of phenomenological characterization is directly transferable to ITER disruption mitigation system design.

██████████ 0.9 plasma-disruption Preprint

ReadSaveConnections

A three-dimensional convolutional neural network trained on atomic-scale potential energy landscapes around defects in tungsten predicts how easily hydrogen atoms migrate through the material, with a mean error of 0.124 eV and computation time of 2.7 milliseconds — more than 23,000 times faster than the conventional quantum-mechanical calculation it replaces. This speed matters because tritium retention in tungsten walls depends on trapping at defects whose behavior requires millions of such migration barrier evaluations to model realistically. Full dataset and code availability are not confirmed, but the architecture is standard and the benchmark method is well-established.

██████████ 0.9 plasma-wall Preprint

ReadSaveConnections

A simulation framework called HISP, built on the open-source FESTIM code, estimates that ten days of ITER deuterium-tritium operation would leave roughly 35 grams of tritium trapped in first-wall and divertor components, with 80% of that inventory residing in co-deposited boron layers on the divertor rather than in tungsten metal. Baking the vessel to elevated temperature was the most effective recovery method, removing 88% of the tungsten-trapped tritium but only 30% of what was in the boron layers. Because tritium is both expensive and subject to safety inventory limits, knowing where it concentrates and how to recover it is operationally critical for any burning plasma device.

██████████ 0.9 plasma-wall Preprint

ReadSaveConnections

A long-standing observation in stellarators — that the ion temperature seems to stop rising past a certain point regardless of heating power — is given a new theoretical explanation here: the magnetic field's inherent aperiodicity causes the turbulent modes responsible for heat loss to become spatially localized rather than global, analogous to Anderson localization of electrons in disordered materials. The authors map the governing equation onto a well-known condensed-matter model (the Aubry-André-Harper equation) and derive a threshold condition that matches W7-X parameters. If confirmed, this could explain why stellarators sometimes achieve better confinement than simple turbulence models predict.

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

ReadSaveConnections

Gyrokinetic simulations using the GENE code show that zonal flows — large-scale plasma rotations that suppress turbulence — are periodically disrupted by bursts of energy flowing back from zonal modes into smaller-scale turbulence, setting a fundamental limit on how long zonal flows can coherently suppress heat transport. Crucially, plasmas with negative triangularity (an alternative cross-sectional shape for a tokamak) exhibit significantly less of this back-transfer, helping explain why negative triangularity configurations experimentally show lower heat losses. The result provides a mechanistic link between plasma shape and confinement quality.

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

ReadSaveConnections

gyaradax is a new gyrokinetic turbulence simulation code written in JAX that runs natively on GPUs and supports automatic differentiation, enabling gradient-based optimization and machine learning workflows that legacy Fortran codes cannot easily support. Benchmarks against the established GKW code show statistical agreement in transport quantities while delivering substantial wall-clock speedup. The code is publicly available on GitHub, and its compatibility with ML frameworks opens a practical path toward training neural network surrogates directly against high-fidelity turbulence simulations.

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

ReadSaveConnections

A graph-based retrieval system for plasma physics literature helps large language models select correct input parameters for gyrokinetic turbulence simulations, outperforming standard retrieval-augmented generation by more than 10% in answer quality and reducing hallucination rates by up to 25%. The approach builds a structured knowledge graph from published plasma physics texts, allowing the model to reason over physical relationships rather than raw document similarity. For a field where setting up gyrokinetic simulations correctly requires specialist knowledge encoded across hundreds of papers, automated parameter assistance could meaningfully lower the barrier for new practitioners.

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

ReadSaveConnections

This theoretical study develops a four-fluid magnetohydrodynamic equilibrium model for a spherical tokamak burning proton-boron-11 fuel, a reaction that produces no neutrons but requires ion temperatures above 100 keV — far hotter than deuterium-tritium fusion. The model accounts separately for thermal ions, suprathermal protons at ~0.5 MeV, relativistic electrons, and bulk electrons, and claims 1% numerical accuracy on macroscopic profiles. While the burning-plasma regime described has no experimental backing and confidence is rated low, the work is relevant as a design-space exploration for aneutronic fusion concepts currently being pursued by several private ventures.

██████████ 0.7 long-confinement Preprint

ReadSaveConnections