An Active Learning Framework for Analog Circuit Multi-objective Customization

Previous research has shown that an analog mixed-signal (AMS) circuit designed by experts to optimally meet one set of specifications, validated via tapeouts and silicon measurements, can be used as a starting point to derive a family of versions of the circuit that efficiently satisfy a wide range of specifications [8, 9, 15]. Their key idea is to retain the netlist topology of the AMS design but to identify alternative combinations of values of circuit parameters. An approach called Analog Mixed-signal Parameter Search Engine (AMPSE) [8] has been developed to enable this search. Thus, this approach reduces the cost of designing AMS circuits by amortizing the extremely high cost of the effort of experts over a family of designs for a wide range of specifications. This parameter search approach faces two main challenges. First, AMS circuit modules have large numbers of parameters, hence the search space is extremely large. Second, circuit simulations required to compute the values of the performance metrics for any given combination of parameter values has extremely high complexity. In this paper, we propose an adaptive sampling method (AS) to identify potential samples, i.e., combinations of parameter values, that enable efficient and effective search of the space. We use a machine learning (ML) based regression model to estimate the values of metrics of the circuit for a given sample. This regression model has dramatically lower complexity compared to circuit simulations. Finally, we develop an active learning approach to reduce the complexity of training the ML-based regression model. In each iteration, we use a mix of uniform and directed sampling to identify parameter value combinations, use the regression model to identify a subset of these samples to simulate, and use the simulation results to retrain the regression model. Our method is more effective and has lower complexity compared with prior methods [12], [17], [5] because it uses a low-complexity yet effective strategy to identify samples for simulation, and uses an adaptive query strategy to balance exploration vs. exploitation. In the experiments, we empirically demonstrated that for analog circuit optimization, at lower complexity, our method discovers better Pareto-optimal designs compared to prior methods. The benefits of our method, relative to prior methods [12], [17], [5], increase as we move from AMS circuits with low simulation complexities to those with higher simulation complexities. For an AMS circuit with very high simulation complexity, our method identifies designs that are superior to the version of the circuit optimized by expert designers.