Enhancing Interpretability of Neural Compact Models: Toward Reliable Device Modeling

Abstract

Neural Compact Models (NCMs) have emerged as a crucial tool to meet the stringent demands of Design-Technology Co-Optimization (DTCO) and to overcome the complexities and prolonged development cycles encountered in traditional compact model creation. Despite their efficiency in simulating electronic devices, a significant barrier to the widespread adoption of NCMs in the industry remains: the lack of interpretability. In the semiconductor sector, where inaccuracies or failures can lead to considerable financial consequences, it is critical to ensure that the model’s predictions are both understandable and reliable. This study aims to enhance the interpretability of NCMs used for I-V and C-V characterizations by clarifying the physical significance of latent vectors. A regularization technique is employed to disentangle features within the latent space, and interpolation is used to visualize and elucidate each dimension’s physical impact. Our approach, which offers interpretable insights into the model’s functionality, seeks to encourage broader implementation of NCMs in the industry, thus accelerating advancements in DTCO.

Keywords

• Neural compact models
• interpretability
• latent vector interpolation
• design-technology co-optimization