284

Papers

10/10

Roadblocks Active

5

Connections

TokaGrad is the first tokamak transport simulator built entirely from differentiable components — plasma equilibrium, transport, heating, L-H transition, and pedestal formation are all linked in a single computational graph that propagates gradients. This means engineers can directly compute how changing a heating waveform or gas fueling rate affects final stored energy, rather than running thousands of trial simulations. The approach could cut the time to design optimal H-mode scenarios from weeks of expert iteration to hours of automated gradient descent.

██████████ 0.9 long-confinement Preprint

ReadSaveConnections

This paper works out the plasma equilibrium design for a 3 GW fusion / 1 GW-electric stellarator based on a scaled Wendelstein 7-X geometry, demonstrating that alpha particle confinement can exceed 85% and neoclassical transport losses can be held below the 0.01 effective-ripple threshold simultaneously. Achieving both at commercial scale in a single equilibrium is non-trivial because optimizing one tends to degrade the other. The design gives the stellarator community a concrete plasma target to build coil, blanket, and engineering design around.

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

ReadSaveConnections

This paper trains a disruption predictor that uses both visible camera images and magnetic coil signals, then compresses its learned knowledge into a lightweight model that uses only magnetic signals — the latter can run in under 5 ms on a real-time control computer. The key insight is that visible images reveal spatial patterns of plasma instability (localized brightening from tearing mode growth) that improve prediction discrimination, and hierarchical distillation transfers this spatial knowledge to the fast model without requiring a camera at inference time. Validated on 640 EAST discharges, the approach directly addresses the tension between richer diagnostics and real-time deployment constraints.

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

ReadSaveConnections

For the first time, coil geometry and mechanical support clamp positions are optimized simultaneously in a stellarator using differentiable finite element stress analysis, achieving a 2.4-fold reduction in peak stress while maintaining magnetic field quality. This matters for high-temperature superconducting (HTS) magnets because the enormous Lorentz forces in high-field coils can degrade or fracture superconducting tape if stress concentrations are not carefully managed. By making the structural mechanics differentiable, the optimizer can trade off magnetic precision against mechanical safety rather than treating them as sequential design steps.

██████████ 0.9 hts-magnets Preprint

ReadSaveConnections

Using linear gyrokinetic theory, this paper shows that electromagnetic corrections to collisionless trapped-electron modes — a major source of energy and particle transport in tokamaks — are inherently small because kinetic and fluid contributions to the parallel electron current nearly cancel each other. This is practically useful because it clarifies when the much cheaper electrostatic approximation is valid in turbulence simulations, potentially reducing the computational cost of transport modeling without sacrificing accuracy. The result applies in the low plasma pressure (low-beta) regime relevant to current and near-term devices.

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

ReadSaveConnections

This study compares two ways of using microwave radar (Doppler backscattering) to measure plasma turbulence spectra in DIII-D, finding that shifting the plasma vertically during a single shot covers more of the poloidal cross-section than varying the antenna angle across multiple shots. Single-shot spectrum coverage is valuable because plasma conditions drift between shots, so comparing different angles measured in different shots introduces systematic errors. The result provides a practical guide for designing turbulence diagnostic campaigns that generate data most useful for validating transport models.

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

ReadSaveConnections

Time-resolved X-ray spectra from krypton tracer atoms in NIF implosions are used to measure electron temperature and density in the compressed hot spot during burn, providing direct experimental constraints on implosion thermodynamics that are otherwise inferred indirectly. The study also finds that tungsten dopants in the carbon ablator measurably alter hot spot conditions and improve implosion performance — a materials-choice result with direct implications for target design. These are among the few direct measurements of hot-spot state during a live fusion implosion, making them important benchmarks for implosion simulation codes.

██████████ 0.8 q-engineering 🔗 18 cited Preprint

ReadSaveConnections

An LSTM-based machine learning model reconstructs full 2D plasma state maps on a 256×256 grid using only three density probe sensors and 50 timesteps of history, demonstrating robustness to sparse and randomly placed sensors. If this generalises to real tokamak or stellarator data, it could dramatically reduce the number of physical diagnostics needed for plasma monitoring and control — lowering cost and improving reliability. The current demonstration is on a particle-in-cell simulation benchmark rather than fusion reactor data, so experimental validation remains the critical next step.

██████████ 0.8 plasma-disruption Preprint

ReadSaveConnections

This theory paper derives guiding-center kinetic equations, force balance with anisotropic pressure, and quasisymmetry conditions from a single variational action principle, showing they are mathematically unified rather than separate assumptions. The practical payoff is a rigorous way to include anisotropic pressure from fusion-born alpha particles in stellarator equilibrium and stability calculations without ad-hoc closures that can introduce unphysical artefacts. The framework also provides a principled starting point for extending gyrokinetic turbulence codes to include anisotropic stress effects important in highly non-axisymmetric geometries.

██████████ 0.7 q-engineering Preprint

ReadSaveConnections

Multiple deep learning architectures — including physics-informed neural networks that incorporate the Grad-Shafranov equation as a training constraint — are trained on 100,760 synthetic equilibrium cases derived from 766 ADITYA-U tokamak discharges to predict plasma shape and pressure profiles faster than conventional equilibrium solvers. Fast equilibrium reconstruction is a prerequisite for real-time plasma control on any tokamak; adding the physics constraint to the loss function modestly but consistently improves accuracy over purely data-driven models. The work also develops an inverse model that estimates the poloidal field coil currents needed to achieve a target plasma shape.

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

ReadSaveConnections