This work presents a complete deep-learning-driven automation flow for designing ultra-low-power subthreshold operational transconductance amplifiers used in electronically tunable gm-C resonator and anti-resonator circuits for speech vowel processing. A multi-task learning neural network is trained as an accurate and extremely fast metamodel (surrogate) of an 8-parameter operational transconductance amplifier topology in 350 nm CMOS process node. Using only 60000 Latin-hypercube-sampled circuit simulations for training, the metamodel predicts DC power consumption, total transistor area, and DC transconductance with average errors of 2.9%, 5.0%, and 7.3%, respectively. The trained model is then coupled with the NSGA-II multi-objective genetic algorithm to instantly generate Pareto-optimal trade-offs between power and area for any designer-specified gm value in the 45–55 nA/V range typical of vowel formant filters. Compared to conventional simulation-in-the-loop NSGA-II optimization, the proposed flow reduces the total design time from 25 h to 7 h, a 3.5 times speedup, and against the standard human design methodology, the acceleration is higher than 20 times. The methodology enables on-the-fly power/area optimization of all operational transconductance amplifiers in sixth-order tunable vowel filters without need for re-design.
Karipidis et al. (Mon,) studied this question.