Machine learning-assisted Real-Time simulation approach for etch and deposition in ultra high aspect ratio (AR >100) structures

Ultra-high aspect ratio features in advanced memory devices present significant challenges for process simulation due to computational burden. While traditional modeling approaches have served adequately for conventional geometries, the emergence of these extreme structures-characterized by their unprecedented depth-to-width ratios and tortuous internal surfaces-has exposed fundamental limitations in our current simulation capabilities, particularly when real-time feedback is essential for process development. We present a machine learning-assisted 3D topography simulation framework that, by circumventing the traditional computational bottlenecks inherent in ultra high-AR modeling, enables near real-time simulation of both etching and deposition processes.

Our framework incorporates two distinct yet complementary modules: and aisotropic etch module and an atomic layer deposition module that meticulously track transient surface reactions of incident species as they navigate the ultra-deep features. The crux of the computationsl challenge lies in accurately determining the flux of neutral species to surface buried deep within these high-AR structures, where traditional view-factor calculations or Monte-carlo ray tracing must account for countless re-emission events, each governed by stochastic processes that, whiel pysically meaningful, impose an exponentially growing computational cost as aspect ratios exceed 100.

Rather than accepting this computational burden as inevitable, we integrated a machine larning surrogate model that learns, through careful training on representative geometries, to directly infer local particle flux from geomteric context alone-effectively encoding the physics of molecular transport into a neutal network that executes in a few seconds rather than minutes or hours. This synergetic coupling of physics-based modeling with data-driven acceleration not only achieves the orders-of-magnitude speedup necessary for interactive simulation but does so without sacrificing the predictive fidelity that process engineers rely upon, thereby transforming what was once a computational bottlenect into a practical tool for optimizaing the increasingly complex fabrication processes demanded by next-generation semiconductor devices.