The rapid evolution of Large Language Models (LLMs) and Generative AI is currently driving a paradigm shift in artificial intelligence innovation. One of the key enablers is high-speed memory, which is why HBM has been attracting attention recently. The need for cloud servers to support digital transformation across a variety of industries requires continued advancements in high-density NAND flash memory. The three main characteristics of memory semiconductors demanded by consumers are high speed, high capacity, and low power. These requirements have been achieved by reducing the size of memory devices. However, as devices shrink in size and stack heights increase, the difficulty of manufacturing technology increases rapidly, making it difficult to launch products. For this reason, computational physics (Technology CAD) is being used more often in the early stages of product development than in the past. In the 2D memory era, the simulation field was primarily device simulation. However in the era of 3D memory, it has expanded to the field of process simulations such as mechanics, thermals, optics, plasma, profile, fluid dynamics, and material simulation to solve more complex development problems, as shown in Fig.1.

While the field of TCAD is expanding from device to process, traditional TCAD simulations have several limitations. First, simulation time poses a serious bottleneck for large-scale domain analysis. Second, there is a lack of physical models that can explain all the physical phenomena that occur during the development of memory semiconductors. Third, multi-physics analysis is necessary, but the appropriate simulation mesh structure for each physical model is different, making it difficult to develop a universal multi-physics solution for memory semiconductor development.

 To overcome these constraints, SK hynix is addressing the limitations of conventional methodologies by integrating TCAD with AI models — a paradigm we define as 'TCAD 3.0' — to significantly enhance productivity in memory semiconductor development. Representative examples include material analysis using Machine Learning Interatomic Potentials (MLP), which has been applied to optimize DRAM capacitor stacks and facilitate pathfinding for new materials, as well as Graph Neural Network (GNN)-based etch simulations designed to derive multi-etch recipes. Beyond these cases, SK hynix is building an AI-augmented TCAD solution that leverages diverse AI models for device and process simulation, material analysis, layout optimization, and physics model calibration, as shown in Fig.2. The company anticipates that the accumulation and integration of these technologies will ultimately enable the realization of Digital Twins and AI Agents for advanced manufacturing as shown in Fig.3