Background: Test patterns are crucial in micro-electronic device patterning because they are fundamental to characterize and model lithography or etch processes.

Problem: Empirical etch models, used to predict wafer pattern shapes, often require extensive pattern characterization to reduce the risk of overfitting with the hope of having good design space coverage. This highlights the critical importance of selecting suitable test patterns. This operation could be facilitated if we had a metric to measure the quality of a test pattern.

Approach: This study utilized post-etch SEM contours from 74 test patterns to train etch models using a neural network, evaluating model performances with the Edge Placement Error (EPE) metric. The idea is to exploit the risk of model instability in neural networks to select and rank the best test pattern candidates based on their model prediction performance. Additionally, it investigates correlations between pattern testing scores and pattern signatures in a multidimensional feature space.

Conclusion: The prediction performance of the machine learning etch model trained from a single pattern is a good metric of the “usefulness” score of the pattern. The scores of the various patterns could be correlated with the pattern feature signatures. Ultimately, this allows selecting or optimizing patterns based on their features only.