Plasma etching profile evolution prediction based on edge-extended long short-term memory network
TL;DR
An edge-extended Long Short-Term Memory network (exLSTM) is proposed for etching profile evolution prediction, combining the exponential gating of scalar LSTM for long-term dependencies and the matrix memory of matrix LSTM for complex spatial modeling.
Abstract
In advanced semiconductor manufacturing, precise control of micro-nano structures via plasma etching is crucial. However, physical simulations are computationally intensive and struggle with rate coefficient determination. While data-driven deep learning offers an alternative, existing methods often suffer from blurry profiles, cumulative errors, and poor modeling of complex patterns. To address these challenges, we propose an edge-extended Long Short-Term Memory network (exLSTM) for etching profile evolution prediction. This model integrates a gradient difference loss term with an Extended LSTM architecture, combining the exponential gating of scalar LSTM for long-term dependencies and the matrix memory of matrix LSTM for complex spatial modeling. The model inputs initial profile images and process parameters to generate temporal evolution sequences. Experimental results demonstrate that exLSTM significantly outperforms the baseline Convolutional LSTM. Specifically, for isolated, dense, and Static Random-Access Memory patterns, the model enhances profile clarity and error suppression, achieving Structural Similarity Index Measure values of 0.9855, 0.9720, and 0.9541 at the 20th frame, i.e., the terminal etching profile at 60 s in a sequence sampled at 3 s intervals, respectively. Furthermore, the model generates full sequences in under 0.015 s on a single Graphics Processing Unit, offering a rapid, high-precision tool for process optimization in advanced node manufacturing.