AI-Assisted Multi-Objective Optimization and Adaptive Control of a Feedback-Enabled 10-nm 10T FinFET SRAM Cell
DOI:
https://doi.org/10.70917/ijcisim-2026-5264Keywords:
10T SRAM, FinFET, 10-nm technology, Feedback-assisted SRAM, Adaptive feedback control, NSGA-II, Reinforcement Learning, Static Noise Margin, Power-Delay Product, PVT variation, SRAM reliabilityAbstract
Designing nanoscale Static Random Access Memory (SRAM) requires balancing power consumption, access speed, stability, and reliability under device and operating variations. Conventional manual transistor sizing relies on repeated circuit simulations and designer experience. This approach becomes inefficient when numerous discrete FinFET sizing options, operating voltages, feedback conditions, and PVT corners must be considered simultaneously. Single-objective optimization is also inadequate because SRAM design requires a balanced compromise among leakage power, dynamic power, propagation delay, Power-Delay Product (PDP), SNM, and reliability. This work presents a feedback-enabled 10T FinFET SRAM at the 10 nm technology node, integrated with a two-stage AI-based optimization framework. Non-dominated Sorting Genetic Algorithm II (NSGA-II) performs design-time multi-objective optimization of transistor configuration, supply voltage, threshold-voltage selection, and feedback strength by considering power, delay, leakage, Power-Delay Product (PDP), and Static Noise Margin (SNM). A reinforcement-learning controller subsequently adjusts the feedback condition according to write, read, hold, and adverse operating states. The SRAM is evaluated using transistor-level HSPICE simulations, Process Voltage Temperature (PVT) analysis, and Monte Carlo variation analysis. Compared with the unoptimized 10T configuration, the NSGA-II–RL design reduces write and read power by approximately 11.62% and 11.95%, respectively, while reducing the corresponding delays by 6.97% and 6.49%. Write and read PDP are reduced by approximately 17.77% and 17.67%, leakage power decreases by 19.09%, and hold SNM improves by 8.65%. The results demonstrate that combining feedback-assisted restoration with multi-objective optimization and adaptive learning provides an effective approach for improving the energy efficiency, speed, stability, and reliability of nanoscale FinFET SRAM.
