Advancing AI-driven computer vision and image segmentation via pattern recognition, GPU-accelerated unsupervised clustering, and edge AI for HPC-scalable big data processing: 85% efficiency gains

Abstract

As data volumes continue to grow, large language models (LLMs) and generative AI (Gen AI) gain popularity, and image resolutions increase across scientific, photography, and industrial domains, the demand for high-performance, GPU-accelerated, and AI-driven image processing has become critical. Emerging applications such as autonomous vehicle perception, medical imaging, satellite image analysis, high-precision manufacturing, robotics, and real-time anomaly detection all require efficient, scalable algorithms to handle massive datasets with high spatial resolution. However, these applications often involve complex image structures, high noise levels, and significant variability, making real-time processing challenging. To address this, we present a GPU-parallelized, cloud-scalable image processing framework that integrates Geometry-Based Machine Learning (GBML) clustering and targeted tuning to significantly reduce computational overhead while preserving critical pattern fidelity. To demonstrate this approach, we chose a challenging, high-precision dataset from advanced nanometer-scale semiconductor manufacturing, where ultra-high-resolution (UHR) Scanning Electron Microscope (SEM) images are critical for precise pattern fidelity assessment and Optical Proximity Correction (OPC) modeling. These images, captured at the physical limits of SEM technology, contain nanometer-scale details essential for device performance but are characterized by extreme noise, focus variations, and high pixel density. This makes them ideal for evaluating the limits of real-time, high-precision image processing algorithms. Traditional OPC models rely on Critical Dimension (CD)-based techniques that extract only localized, linear cross-sections from SEM images, failing to capture the full spatial complexity required for accurate contour modeling. As industry shifts toward contour-based OPC, which provides a holistic representation of lithographic structures, efficient and scalable contour extraction has become a critical requirement. However, extracting contours from UHR SEM images presents significant challenges, including high computational costs, process-induced variability, and the need to preserve metrology-critical pattern fidelity. In this work, we propose a novel methodology that integrates GPU-accelerated, AI-driven pattern recognition with GBML clustering and targeted tuning focused on regions of interest (ROI). Implemented within Calibre SEMSuite™, this approach optimizes real-time contour extraction by clustering structurally similar SEM images, significantly reducing computational overhead without sacrificing precision. By avoiding redundant tuning across dissimilar patterns, our approach ensures high-fidelity contour accuracy while minimizing processing time. Edge-optimized GPU processing enables fast local execution, while HPC-enabled cloud scalability supports full-resolution contour extractions across large-scale SEM datasets. Benchmark results demonstrate that GBML clustering alone achieves a 72% improvement in computational efficiency, while the combination of GBML clustering and targeted tuning achieves 85% efficiency gains. In contrast, applying targeted tuning to randomly grouped images results in lower improvement percentages, confirming the importance of structural similarity in optimization. This work demonstrates that AI-driven contour modeling not only preserves critical detail but also scales effectively for next-generation AI and high-precision applications.