The technique at the heart of modern overlay control is diffraction-based overlay, or DBO. Instead of imaging the wafer, a DBO tool shines light onto a specially designed grating target and measures how the light diffracts into positive and negative first orders. In the ideal case, the difference between the two diffraction signals is strictly proportional to the overlay shift between layers. This approach has largely displaced older image-based metrology because it offers superior precision and better resilience to tool-induced shifts. But the clean proportionality assumption breaks down in the real world. Processes such as chemical mechanical polishing and plasma etching leave their fingerprints on the grating targets: sidewalls that lean at different angles on the left and right, rounded corners, and imbalanced critical dimensions. These geometric imperfections generate spurious asymmetry signals that can masquerade as genuine overlay shifts, producing what metrologists call process-induced shift errors that quietly corrupt the measurement.

Hsieh and Yang frame this problem in an illuminating way: process-induced asymmetry is fundamentally an optical leakage problem. Structural perturbations of the grating generate diffraction asymmetry that has nothing to do with overlay, and this nuisance signal leaks into the measurement channel that is supposed to carry only overlay information. Their proposed remedy is to stop reducing the diffraction data to a single scalar asymmetry value at one wavelength, and instead capture the full broadband spectral response of the target. By recording the positive and negative first-order diffraction efficiencies for both TE and TM polarizations at 61 wavelengths spanning 400 to 700 nanometers, each measurement becomes a 244-dimensional vector. The physical premise is that wavelength-dependent diffraction acts as a fingerprint of the grating’s state, one rich enough to distinguish true overlay from the spectral distortions introduced by imperfect processing.

To test this premise rigorously, the researchers built a simulation framework based on the finite-difference time-domain method, using a nonuniform mesh with steps as fine as 0.25 nanometers and, in the critical asymmetric target region, 0.10 nanometers. They designed a hierarchy of three evaluation phases of increasing structural complexity. Phase I isolates a single nuisance mechanism: asymmetric sidewall angles, with the left and right sidewall angles of the bottom grating independently randomized between 80 and 90 degrees. Phase II adds top-corner rounding and global film-thickness variation on top of the sidewall asymmetry. Phase III moves to an advanced PolyFin spacer architecture, mimicking the multilayer dielectric environment of modern gate structures, with overlay plus eight independently varied nuisance variables including left and right poly critical dimensions, sidewall angles, spacer corner radii, and two thickness variations. The datasets grew from 932 samples in Phase I to 2,394 in Phase III, each sample generated by a separate full electromagnetic simulation.

On top of these simulated spectra, the authors posed the inverse problem: given only the raw 244-feature spectrum, with no geometry metadata, estimate the overlay. Crucially, they did not bet everything on deep learning. They compared linear regression, Ridge regression, RBF support vector regression, XGBoost, and a four-hidden-layer deep neural network, all trained on identical raw inputs with identical data partitions and validation-only model selection. This design deliberately separates two questions that are often conflated: how much overlay information does the broadband optical response actually contain, and how much does the choice of estimator matter? The results delivered a surprising verdict. In the noise-free Phase I dataset, plain linear regression achieved a root-mean-square error of just 0.0015 nanometers, meaning overlay was almost perfectly linearly decodable from the broadband spectrum. In the most complex Phase III scenario, linear regression and Ridge still reached 0.229 and 0.230 nanometers respectively, beating the DNN’s 0.374 nanometers and far outperforming XGBoost at 1.564 nanometers.

The contrast with conventional narrowband approaches was stark. A fixed single-wavelength reference at 550 nanometers collapsed to errors of roughly 8 to 9 nanometers under compound process imperfections, and even the best single wavelength managed only 5.2 nanometers in the PolyFin case. Multi-wavelength inputs changed the picture entirely, and a Ridge regression on paired-target broadband features reached 0.299 nanometers in Phase III. The authors are careful about what this does and does not prove. The deep network is a competitive nonlinear estimator, but the study explicitly does not establish universal DNN superiority; rather, it shows that broadband spectral observability, rather than model complexity alone, drives much of the performance. No inverse model, however expressive, can recover overlay information that is absent or locally confounded in the measured channels.

The team backed the headline numbers with an unusually thorough battery of robustness analyses. Empirical spectral diagnostics showed that the wavelength carrying the strongest overlay association shifts with the target: 485 nanometers in Phase I, 590 in Phase II, and 690 in Phase III. At individual wavelengths, the overlay response can become nearly parallel to a nuisance response, creating local ambiguity, but across the full 244-feature space the largest global overlay-geometry alignment in Phase III was a moderate cosine of 0.339, and in Phase I the overlay and sidewall-asymmetry responses were nearly orthogonal. Singular-value analysis revealed that the spectra are highly redundant, with only 4, 5, and 14 components explaining 95 percent of the training variance across the three phases, though the very large condition numbers warn that unregularized inversion can amplify small perturbations dramatically.

That warning proved prophetic in the noise stress tests. When trained on noise-free data and then confronted with test spectra perturbed by independent relative noise, the models diverged sharply. Unregularized linear regression, which had appeared almost magical on clean data, exploded to errors of tens of nanometers at just 1 percent noise, while Ridge and the DNN degraded far more gracefully; the Phase III DNN moved only from 0.374 to 0.406 nanometers at 1 percent noise. A blocked holdout test, in which the 20 percent of samples closest to the boundary of the simulated process space were withheld entirely, showed all models degrading by 35 to 68 percent relative to in-domain performance, a reminder that interpolation within a known process window does not automatically transfer to unseen conditions. Bootstrap resampling and repeated data partitions confirmed that the Ridge-versus-DNN difference was statistically robust for the fixed test set, while wavelength ablations showed that five evenly spaced combined-polarization wavelengths already deliver sub-nanometer accuracy, though matching the full-spectrum result to within 10 percent required all 61 labels.

The authors are candid that this is a simulation-based study with defined limits: the two-dimensional model excludes finite mark size, line-edge roughness, wafer-scale nonuniformity, and tool drift, and the noise model is a generic sensitivity test rather than a calibrated instrument specification. The sub-nanometer errors therefore establish feasibility within the evaluated simulations, not validated production-line accuracy. Yet the practical pathway they sketch is concrete. They propose a hybrid digital-twin workflow in which a convergence-verified FDTD library pretrains an inverse estimator, a deliberately selected set of measured calibration data adapts it to the real tool, active learning targets high-uncertainty regions, and continuous monitoring of residuals and spectral distance triggers fallback to reference metrology when drift or out-of-distribution conditions appear. They also point out that target pitch shapes which wavelengths carry overlay information, opening the door to co-designing grating targets and measurement bandwidths.

For an industry where every fraction of a nanometer of edge-placement error translates directly into yield and performance, the message of this work is quietly radical: the information needed to defeat process-induced overlay errors may already be present in the full spectrum of light bouncing off the wafer, waiting to be read properly. The finding that simple, well-regularized linear models can rival deep networks when given rich broadband inputs is a useful corrective to the assumption that harder problems always demand bigger models. As chipmakers push toward gate-all-around transistors and beyond, the combination of full-spectrum diffraction fingerprints, physics-guided simulation, and disciplined statistical validation may prove to be the microscope that keeps the nanoscale world in focus.

Subject of Research: Broadband diffraction-based overlay metrology for semiconductor lithography under process-induced structural asymmetry

Article Title: Broadband spectral metrology for robust diffraction-based overlay estimation under process-induced asymmetry

Article References: Hsieh, H.-C., & Yang, C.-S. (2026). Broadband spectral metrology for robust diffraction-based overlay estimation under process-induced asymmetry. Results in Optics, 25, Article 101156. https://doi.org/10.1016/j.rio.2026.101156

Image Credits: AI Generated

DOI: 10.1016/j.rio.2026.101156

Keywords: overlay metrology, diffraction-based overlay, semiconductor manufacturing, broadband spectroscopy, FDTD simulation, process-induced shift, deep learning, ridge regression, lithography, nanometrology, digital twin, PolyFin spacer