As semiconductor manufacturing advances toward nanoscale technology nodes, plasma etching has become increasingly critical for fabricating high-aspect-ratio (HAR) structures. In these structures, understanding ion transport dynamics is essential for predicting etch profiles and optimizing process conditions. Atomic-scale simulations can accurately capture ion-surface interactions, but simulating entire nanoscale trenches at full atomic resolution is computationally prohibitive. Monte Carlo (MC) methods offer computational efficiency but require extensive pre-computation for event table construction and fail to resolve the atomistic-level collision dynamics necessary for accurate predictions. To overcome these limitations, we developed an AI-hybrid multiscale simulation framework that integrates the physical accuracy of molecular dynamics (MD) with the computational efficiency of MC methods. First, we performed one million MD simulations of single Ar ion bombardment on Si substrates using LAMMPS, systematically varying incident energies (0.1–100 eV) and polar angles (0–89°) [1]. The resulting energy-angle distributions were used to train a Mixture Density Network (MDN) that predicts probability density functions for ion reflection and implantation events. Unlike standard neural networks, the MDN captures the stochastic nature of ion-surface interactions by generating complete probability distributions [2]. The trained MDN model replaces computationally expensive MD calculations within our MC framework for simulating ion trajectories in 2D trench geometries. Our simulations show that aspect ratio significantly restricts ion penetration depth, with energy deposition concentrated near trench entrances and bottoms, providing quantitative insights into how trench geometry affects etching uniformity. Additionally, we developed an LLM-based autonomous agent system that automates the entire simulation workflow. This system accepts user-specified ion species, substrate compositions, and structural parameters (including images), then automatically constructs MD frameworks, executes simulations, trains MDN models, and generates results. This capability enables rapid adaptation to different material systems, making the framework a practical tool for optimizing plasma etching processes in next-generation semiconductor manufacturing.