New MOSFET modeling algorithms and their use in CAD of analog IC building blocks

Analog integrated circuit (IC) design has undergone several technical advancements following Moore's law, and tends to become extremely challenging with the continued downscaling of the devices and supply voltages. However, not many sophisticated and detailed design tools are available to aid analog designers in exploiting the complete potential of these technical advancements. Most of the available commercial and in-house design tools model the basic building blocks (e.g. transistor) and employ these device models to predict the performance of a complete circuit/system. As such, accuracies of these device models are crucial in order to develop efficient design tools. Typically, accurate models could be complex, while simple models could be inaccurate. As such, new modeling algorithms leading to simple yet accurate device models and satisfactory design tools continued to be in great demand.;Finally, device models developed using the aforementioned modeling algorithms are employed to build an accurate and extendable computer aided design (CAD) tool for the design of analog IC building blocks (e.g. current sources/sinks, single stage amplifiers, simple and cascode current mirrors, voltage divider, differential amplifier, and three stage operational amplifier).;In this thesis, neural networks that offer advantages like simple calculations and a wide spectrum of applications, are employed for the modeling purpose. Firstly, new modeling algorithms based on binning concepts that offer accurate device modeling over wider input parameter space of the problems that have outputs highly non-linear to one of the inputs are proposed. Both single and multi-dimensional modeling algorithms are developed, and illustrated through device modeling examples. A new neural modeling approach based on a correction model is then introduced for the first time to develop accurate device level models for highly non-linear input-output behaviours that are difficult/impossible to model with simple structures. The proposed approach simplifies the modeling process for a novice/inexperienced designer as it eliminates the need of in depth understanding of the neural network concepts by virtue of using well known simple 3-layer MLP networks. MOSFET modeling example confirms that the approach leads to accurate neural models while keeping the model structure simple.