Efficient Memory Circuits Yield Analysis and Optimization Framework via Meta-Learning

Yield optimization is a computationally intensive process that requires repeated yield estimation. In this paper, we propose a yield-driven optimization process based on meta-model yield analysis for memory circuits. Choosing the Artificial Neural Network (ANN) as surrogate model, we propose a model generation strategy to learn better initialization of network parameters via meta-learning. We improve the Differential Evolution (DE) algorithm to guide the optimization process and fine-tune the yield surrogate model in each iteration with limited simulation overheads. The experimental results validated on the memory circuits show that the proposed method can achieve final results with higher yields and performs high accuracy and robustness compared to state-of-the-art methods.