Generative AI Driven Multi-Dimensional Data Modeling for Variability Aware Synthetic Electrical Characteristics Generation in Complementary Field-Effect Transistors

• Generative AI framework for variability-aware synthetic semiconductor data generation. • Hybrid LSTM-CNN conditional GAN captures complex device variability patterns. • Multi-dimensional data integrates process and electrical information. • High-fidelity synthetic device characteristics with 0.5% error rate. • Reduces reliance on computationally expensive 3D device simulations. Generative artificial intelligence (Gen AI) has emerged as a transformative paradigm for addressing data scarcity and computational challenges in semiconductor device research. Modern semiconductor design depends on the availability of high-quality large-scale device data. This requirement is challenging to meet because three-dimensional numerical simulations and experimental measurements are computationally intensive and time-consuming. To address this challenge, this study proposes a generative AI framework for efficient and scalable synthetic data generation to analyze process variation effects in advanced nanoscale devices. The approach employs a hybrid conditional GAN architecture with an LSTM-based generator to capture sequential dependencies and a 1D-CNN discriminator to enforce local structural realism in the generated data. The proposed adversarial framework enables high-fidelity generation of electrical characteristics for stacked nanosheet CFETs under multi-dimensional process variations. Experimental results show strong agreement between synthetic and simulation data, validated through error metrics, adversarial loss, and distribution analysis using Principal Component Analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE). Across all evaluation metrics, the framework achieves a low average error rate of approximately 0.5%, confirming the realism and reliability of the synthetic data generated. We have also performed a comparative analysis between conventional homogeneous GAN architectures and the proposed hybrid model demonstrates the superior performance of the hybrid design, validated through monotonicity evaluation, nearest-neighbor distance, and regression metrics. The proposed framework provides an efficient and reliable data-driven solution for variability analysis and design exploration in next-generation semiconductor devices.