Metrology and testing is the measurement layer that decides whether a fabricated material, structure or device matches intent, and the current wave of work is less about new sensors than about extracting far more information per measurement through autonomous acquisition planning, phase-resolved probes and computational inversion.
Summary
Metrology and testing covers the instruments, protocols and inversion mathematics used to determine what a manufactured object actually is: its composition, strain state, defect population, electronic structure, thermal behaviour and, in security contexts, its provenance. In semiconductor and advanced-materials manufacturing it is a gating function rather than a value-add: a process cannot be tuned faster than the loop that measures its output, and a failure mode that cannot be resolved cannot be engineered away.
The recent literature clusters into four families. First, scanning-probe and electron probes, where the frontier is acquisition strategy rather than resolution: a live closed-loop workflow using a multitask Gaussian process now selects both the next location and the next measurement protocol on an automated large-sample atomic force microscope, demonstrated on a composition-spread AlScN wafer using tapping mode and Dual AC Resonance Tracking. Related electron-optical work extends transmission electron microscopy to compensated magnetic order via electron magnetic linear dichroism, reconstructing the spin axis from momentum-resolved electron energy-loss spectra, and corrects the normalisation used in electron correlation microscopy, where importing X-ray photon correlation spectroscopy conventions unchanged produces artificial anticorrelations and biased relaxation times.
Second, diffraction and scattering. Dark-field X-ray microscopy in reflection geometry has been applied to a 7 micrometre thick (111) HgCdTe epilayer on CdZnTe, separating elastic strain from lattice tilt at individual threading dislocations, which previously were accessible only through topography that conflates the two. Ion channeling and backscattering interpretation has been put on a Monte Carlo footing that distinguishes point from extended defects across analysing-beam energies. Third, direct electrical probes of electronic structure: a lock-in electric-field penetration technique maps the effective density of states in indium-based oxide thin-film transistors, showing exponential suppression of the effective density of states with channel length and giving a microscopic origin for a localisation-induced threshold-voltage roll-off.
Fourth, and cutting across all of the above, the computational half of metrology: forward models and surrogates fast enough to sit inside the measurement loop. Examples include an accuracy-preserving Debye scattering engine that computes a high-resolution total scattering profile for 90 million atoms in a (0.1 micrometre) cube in minutes on distributed CPUs, a machine-learning charge density route that predicts defect formation energies in 360-atom supercells to below 0.05 eV mean absolute error from only 96 training supercells of 16 to 96 atoms, where interatomic potentials trained on the same data err by more than 1 eV, and an operator-learning thermal simulator for 3D ICs reported at 842 times the speed of finite element methods. The parameters that decide the field are therefore not resolution alone but the triangle of resolution, throughput and destructiveness, whether the measured observable maps uniquely onto a device parameter, and whether the inversion is trustworthy enough to act on automatically.
Evidence base
- 11 Aug 2026: a live closed-loop multitask Gaussian process on an automated large-sample AFM selected both the next measurement location and the next protocol, demonstrated on a composition-spread AlScN wafer with tapping-mode and DART measurements.
- 11 Aug 2026: dark-field X-ray microscopy in reflection geometry imaged substrate threading dislocations and the layer features they nucleate in a single frame in a 7 micrometre thick (111) HgCdTe epilayer on CdZnTe, separating elastic strain from lattice tilt that topography conflates.
- 11 Aug 2026: lock-in electric-field penetration mapping of indium-based oxide thin-film transistors showed exponential suppression of the effective density of states with channel length and identified a localisation-induced threshold-voltage roll-off mechanism.
- 11 Aug 2026: a machine-learning charge density model trained on 96 supercells of 16 to 96 atoms predicted defect formation energies in 360-atom supercells to below 0.05 eV mean absolute error, where interatomic potentials on the same data erred by more than 1 eV.
- 11 Aug 2026: an accuracy-preserving Debye scattering engine computed a high-resolution total scattering profile for 90 million atoms in a (0.1 micrometre) cube in minutes on a distributed-memory CPU platform.
- 11 Aug 2026: a segmentation model trained exclusively on GAN-generated synthetic SEM images transferred to real images and outperformed a baseline trained on the limited real dataset, while reproducing none of the original proprietary routing.
- 11 Aug 2026: an opto-electronic oscillator was demonstrated with phase stability approaching the standard quantum limit and then quantum-engineered beyond it, the first demonstration of quantum enhancement in a feedback oscillator.
Open questions
- Can the strain and tilt separation achieved by reflection-geometry dark-field X-ray microscopy on HgCdTe epilayers be reproduced on a laboratory-scale source, or does it remain synchrotron-bound and therefore a diagnostic rather than a process-control tool?
- Does the sim-to-real transfer of synthetic SEM training data hold across different foundries, nodes and imaging conditions, or only within the design family the StyleGAN was fitted to?
- How do the accuracy advantages of charge-density learning and neural-operator surrogates behave out of distribution, given that both are validated against in-family targets?
- What validation regime prevents degenerate observables from being trusted in an autonomous loop, given that quantum capacitance alone cannot distinguish topological from trivial states?
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.