AI-guided plasma simulations separate stronger heating from brighter X-rays

A new preprint by Bowen Zhu and Jian Wu presents Simjecture, an open-source framework designed to organize computational physics around attempts to disprove explicit conjectures. Posted on November 1, 2026, the study combines a research-workflow demonstration with simulations of a Z-pinch: a plasma compressed by the magnetic field associated with a strong electric current. Its findings concern a computational model, not a new laboratory measurement.

The central physical puzzle is why a wire-array Z-pinch can radiate more X-ray energy than the imploding material’s maximum instantaneous bulk kinetic energy. As the authors explain, that kinetic energy is only one stage in a continuously driven energy-transfer process. Comparing cumulative radiation with a momentary kinetic-energy inventory does not account for the full magnetic reservoir or ongoing energy input.

Testing conjectures, not just generating results

Simjecture asks humans to define the physical question, scope and resource budget. A director agent coordinates the campaign, while a worker configures simulations, develops diagnostics and examines results. Separate methods and claim reviewers evaluate numerical reliability and scientific conclusions.

The distinction matters because an apparent counterexample can reflect an inadequate calculation rather than unexpected physics. Under the framework’s rules, a prediction is falsified only when the simulation’s entire qualified numerical measurement range falls outside it. Overlapping ranges leave the question undecided.

Revised hypotheses must also make commitments before new tests run. Simply widening a prediction to accommodate previous results is not considered an informative repair.

In a preliminary two-dimensional magnetic-reconnection study, the system found a qualified counterexample to a specified reconnection-rate scaling band. Two proposed replacement models then failed their prospective calibration checks. The authors report those failures rather than claiming a completed theoretical explanation. This result challenges the tested conjecture within its defined setup, not reconnection theory generally.

More dissipation, less radiation

The main investigation ran for 82.6 hours and used three-dimensional radiation-magnetohydrodynamic simulations in FLASH. The model was anchored to an eight-wire aluminum array on the Zebra pulsed-power generator, but began with a prescribed, already ablated plasma distribution. It therefore omitted the early evolution of cold wires and used an idealized constant one-megaampere drive.

One fine-mesh calculation emitted 3.091 kilojoules during a specified stagnation interval, compared with a tagged-load kinetic reference of 0.460 kilojoules. Radiation was thus 6.72 times that reference. The stored magnetic energy fell by 2.452 kilojoules over the interval, making it the largest declining stored-energy component. A further accounting term combined boundary energy exchange with any numerical residual; it was not an independently resolved measurement of every incoming energy flux.

The campaign’s starting conjecture concerned whether non-axisymmetric magnetic reconnection directly supplied at most 10% of the positive radiation excess. Reconstructed resistive heating reached 12.55% of that excess in the selected fine-mesh comparison. Crucially, the authors do not identify this as a measurement of reconnection’s contribution. Resistive heating also includes diffusion without reconnection, while reconnection can energize bulk motion. Dedicated magnetic-topology diagnostics and finer mesh convergence remain necessary to isolate its share.

A separate comparison produced a counterintuitive trend. Doubling an imposed helical displacement increased resistive conversion but weakened transverse compression. Over 32 nanoseconds, total radiation fell by about 5.3%, while more magnetic and thermal energy remained stored. Within these simulations, stronger dissipation did not automatically make the plasma brighter.

Timing changes the comparison

Broadening the surrounding trailing-plasma distribution produced a different response. In a matched coarse-grid pair, the broader profile diverted current away from the central region early and delayed the later radiation maximum by about five nanoseconds.

At 32 nanoseconds, it had emitted roughly 4% less total radiation. By 44 nanoseconds, subsequent emission had recovered 60% of that earlier deficit, narrowing the cumulative difference to 1.37%. Its cumulative emission above one kiloelectronvolt nevertheless remained 6.21% lower. Total energy output and spectral hardness therefore did not recover identically.

These comparisons have explicit limits: finite observation windows, idealized initial conditions and resolution sensitivity. The trailing-width result came from coarse simulations, and other comparisons showed appreciable mesh effects, particularly in hard X-rays.

The autonomy was also bounded. Four human steering interventions redirected the inquiry, and the authors report that unguided agents tended toward simpler proxies or premature finite-window conclusions. The study offers a documented example of human-guided computational investigation—not a demonstration that AI independently solved the Z-pinch energy problem.