This Week in Nuclear Fusion

Stellarator engineering took a major step forward with a detailed 1 GW-electric fixed-boundary equilibrium design demonstrating that alpha confinement and neoclassical losses can both meet power-plant thresholds simultaneously. On the fuel cycle front, a critical modeling assumption underpinning tritium self-sufficiency was overturned: edge puffing, not core fuelling, dominates particle injection in detached operation by roughly an order of magnitude, forcing a reassessment of tritium inventories. Meanwhile, machine learning methodology crossed into plasma control with TokaGrad, the first fully differentiable tokamak transport simulator. Taken together, this week's papers tighten the engineering requirements for both magnetic confinement and tritium breeding while offering a new computational tool to optimize around them. The field is visibly shifting from proof-of-principle physics toward systems-level power plant design.

Top 3 Papers

1. Fixed-boundary equilibrium for a 1 GW-electric stellarator power plant (GIGA) The GIGA design achieves >85% alpha power confinement alongside a neoclassical effective ripple below 0.01—well beneath typical stellarator benchmarks—within a single self-consistent 3 GW fusion power equilibrium. This is one of the most complete stellarator reactor equilibria published, bridging the gap between optimized configurations and credible power plant geometries.

2. Plasma–tritium fuel cycle coupling through matter injection and particle exhaust Fuel puffing rates in detached operation outpace core fuelling by roughly 10× across present-day tokamaks and next-step stellarators, a regime that legacy tritium fuel cycle (TFC) models systematically ignored. The Direct Internal Recycling assumption of near-equal D:T composition breaks down once puffing's tritium fraction is correctly accounted for, implying that tritium inventories for ITER-class and beyond machines have been underestimated.

3. TokaGrad: End-to-end differentiable tokamak simulator for full-scenario optimization TokaGrad propagates Jacobians across a coupled chain of equilibrium, transport, heating, L-H transition, and pedestal models, enabling true gradient-based optimization of entire discharge waveforms for the first time. This replaces expensive black-box parameter searches and opens the door to jointly optimizing reactor design variables and actuator control schedules within a single differentiable pipeline.

Connection of the Week

Nuclear Fusion ↔ Machine Learning / Differentiable Programming

Bridge logic: TokaGrad is, at its mathematical core, an application of backpropagation—the same chain-rule Jacobian propagation that trains deep neural networks—applied to a chain of coupled physics solvers instead of a chain of neural network layers. The "differentiable everything" paradigm pioneered in ML (JAX, PyTorch autograd) made it computationally tractable to attach automatic differentiation to arbitrary computational graphs. Plasma physicists have now imported this machinery wholesale: a tokamak discharge scenario becomes a deep computational graph, and gradient descent replaces the evolutionary algorithms or manual iteration that previously drove scenario optimization. The implication extends beyond convenience—it means reactor design variables and real-time control parameters can, in principle, be co-optimized in the same loss landscape, collapsing what were previously two separate engineering workflows into one.

Want More?

This digest covers 3 of the 702 papers published in fusion this week. Get daily full digests with all connections, ToT reasoning chains, and roadblock tracking across the entire literature. Upgrade to Pro ($9/mo).