Assessing test pattern quality for etch modeling using neural network prediction stability

BackgroundTest patterns play a critical role in microelectronic device patterning, as they are essential for characterizing and modeling lithography and etch processes.AimEmpirical etch models, which predict wafer pattern shapes, often require extensive pattern characterization to mitigate overfitting and ensure adequate design space coverage. This underscores the importance of selecting suitable test patterns—a task that could be greatly simplified by introducing a metric to quantify pattern quality.ApproachWe use post-etch scanning electron microscopy (SEM) contours from 74 test patterns to train 74 individual etch models with a neural network. Model performance is evaluated using the edge placement error metric. The approach leverages the inherent instability of neural networks as a means to rank test pattern candidates based on their predictive performance.ResultsWe demonstrated that, for each individual pattern, the corresponding model performance is correlated with its pattern signature and can be reliably fitted within a multidimensional feature space.ConclusionsThe prediction performance of a machine-learning-based etch model trained on a single pattern serves as a robust metric for assessing the “usefulness” of that pattern. Furthermore, these scores can be correlated with pattern feature signatures, enabling pattern selection or optimization based solely on feature data following a one-time SEM-based calibration.