Machine Learning Ocd

Machine learning-based optical critical dimension measurement (ML-OCD) replaces the traditional library-matching or regression math behind scatterometry with a trained machine learning model, letting a measured optical spectrum get converted directly into a critical dimension, height, or sidewall angle in microseconds rather than through a slower, more conventional lookup or fitting process. It's the same underlying optical measurement — scatterometry, inferring structure from how light reflects off it — with the slowest, most brittle part of the analysis replaced by a model that learned the relationship between spectra and dimensions instead of computing it from first principles every time.

Conventional scatterometry works by matching a measured optical spectrum against a large pre-computed library of simulated spectra, each corresponding to a specific combination of structural parameters — and that matching process is exactly what ML-OCD speeds up. Traditional library matching searches through potentially millions of simulated spectrum-to-parameter pairs to find the best match, or alternatively runs an iterative regression that repeatedly simulates candidate structures until the simulated spectrum converges on the measured one. Both approaches work, but both require real computation at the moment of measurement, which limits how fast a production metrology tool can report results.

ML-OCD moves nearly all of that computational burden to a training phase that happens once, offline, well before any production measurement ever needs results. Rigorous coupled-wave analysis (RCWA) simulations generate a large synthetic training dataset, pairing known structural parameters with the optical spectra those parameters would physically produce. A machine learning model — a neural network, a Gaussian process, or another suitable architecture — is then trained on this dataset to learn the underlying mapping from spectrum to parameters directly, rather than needing to search or iterate through that mapping at measurement time.

Once trained, inference — predicting structural parameters from a newly measured spectrum — takes microseconds, because the model is simply evaluating a learned function rather than searching a library or iterating toward convergence. This speed difference matters enormously in a high-volume production fab, where metrology tools need to keep pace with wafers moving through the line continuously; a measurement technique that takes meaningfully longer per wafer than the production line's own throughput becomes a bottleneck in its own right, no matter how accurate it is. ML-OCD's near-instant inference removes measurement speed as a constraint, letting metrology keep pace with production without forcing a tradeoff between measurement thoroughness and line speed.

ML-OCD models also tend to handle real-world measurement imperfections — noise, systematic tool error, model simplifications — more gracefully than exact library-matching or regression approaches, because the model learned from training data that itself included these imperfections. A library-matching approach assumes the measured spectrum closely resembles one of its pre-computed library entries, and performs poorly when real measurement noise or systematic deviations push the actual spectrum away from any clean library match. A well-trained ML model, especially one trained with realistic noise and variation built into its synthetic training data, can generalize across exactly this kind of imperfection rather than failing outright when a measurement doesn't match a library entry cleanly.

ML-OCD is particularly valuable for geometrically complex structures — gate-all-around transistors, complementary FET (CFET) stacks — where traditional library matching or regression approaches start to strain against their own computational limits. As device structures have grown more three-dimensional and parameter-rich, the library-matching approach's search space grows correspondingly, and iterative regression's per-measurement simulation cost grows along with it — both approaches scale poorly as structural complexity increases. An ML model, once trained on a sufficiently representative dataset covering that complex structure's parameter space, performs inference at the same microsecond speed regardless of how geometrically complex the underlying structure actually is, which is exactly why ML-OCD adoption has tracked closely with the industry's broader shift toward increasingly complex three-dimensional transistor architectures.