Fully Automated, Scalable, and Robust Generation of Testbenches for Simulation and Optimization of Amplifier Circuits

Abstract

As in other areas, unattended data mining and generation is crucial for the successful application of artificial intelligence in analog design. Unfortunately, analog design knowledge belongs among the most closely guarded domains in the tech industry. This fragmentation and general scarcity poses a serious challenge to keep up with areas that are more accessible to machine learning. In the absence of publicly available, high-quality data, reliable methods for data generation are of huge value. In this work, we describe a robust process for automatically generating testbenches for sizing and evaluation of analog integrated amplifiers, targeting automatic data generation and evaluation of existing sizing data and results without human interaction. While parts of this process are straightforward for the analog designer, the amount of data needed prevents a manual approach. On the other hand, finding general methods for automation is often difficult and requires thorough consideration regarding their validity. In this work, we pay special attention to so far uninvestigated problems in the recognition of important circuit parts, such as detection of inverting and non-inverting input terminals and symmetry. The methods are thoroughly tested and accompanied by mathematical proofs of their general validity. Our test set for verification comprises a massive ~500 operational amplifiers and the complete sizing approach, including optimization, is evaluated with a complete Pareto optimal front for ~100 of those amplifier topologies using the open-source SkyWater-130 PDK. As the complete setup process is automated, this leads to a roughly estimated 99.89% reduction of manual effort.