Witte, Jermann, and colleagues at multiple institutions achieved 25,000 reconstructed volumes per second of laser melt pools in alumina using rotation-enabled X-ray Multi-Projection Imaging at the MAX IV synchrotron. The technique captures three angularly resolved X-ray projections per time step, then uses deep-learning reconstruction to produce full 3D volumes. The critical innovation was decoupling temporal resolution from sample rotation speed — prior tomographic methods required rotating the sample, limiting frame rates to hundreds of hertz. By fixing the sample and using multiple projection angles simultaneously, the team broke through to kilohertz-scale volumetric imaging.
The through-claim is that the observability of a process determines whether it can be controlled. Laser powder bed fusion has been used industrially for years, and the physics of melt-pool formation is reasonably well understood in theory. But melt pools evolve on millisecond timescales in three dimensions, and until this work, no imaging technique could capture their full geometry fast enough to observe defect formation as it happens. The defects — keyholes, pores, incomplete fusion — were diagnosed after the fact from finished parts. The 4D movie does not change the physics; it moves the diagnosis from post-mortem to real-time, which is the prerequisite for closed-loop control.
This is the pattern of observability bottlenecks: systems where the theory is adequate but the measurement is not. Turbulence modeling in weather prediction: the equations are known but the initial conditions are undersampled. Neuronal circuit mapping: the biophysics is understood but simultaneous recording from enough neurons is only now becoming possible. In each case, the science was waiting not for a better theory but for a faster camera.
(arXiv:2603.14391)