JÄNES: A NAS Framework for ML-based EDA Applications

While Machine Learning is increasingly adopted for a spectrum of complex and time-consuming Electronic Design Automation (EDA) tasks, the efficiency of ML-based solutions varies. The targeted low-cost predictions in the EDA flows may require Neural Network (NN) architectures tailored to the task class and a combination of features and specifics of the design under analysis. Neural Architecture Search (NAS) frameworks automate the search process and differ by the types of the supported NNs and the algorithms used at the backend. This paper proposes a novel jumping and Bayesian optimization-based NAS framework focused on the Multi-Layer Perceptron (MLP) NNs that are widely applied for the ML-based EDA tasks. The efficiency of the proposed framework is evaluated on an industry-scale case study for wafer-level manufacturing test die re-test prediction, achieving 82 % accuracy. The experimental results demonstrate a 6 times faster search along with higher efficiency of the resulting NN architecture than a state-of-the-art NAS framework.