Complex network based machine learning method for predicting circuit timing

Tingyuan Nie, Mingzhi Zhao, Pengfei Liu, Zhenhao Wang,
Complex network based machine learning method for predicting circuit timing,
Applied Soft Computing,
2025,
114100,
ISSN 1568-4946,
https://doi.org/10.1016/j.asoc.2025.114100.
(https://www.sciencedirect.com/science/article/pii/S1568494625014139)
Abstract: Due to the high cost of timing signoff in the routing stage of VLSI (very large-scale integration), timing analysis in the placement stage is usually used to assist the implementation. Still, uncertainty exists due to the lack of detailed wiring information. Considering the close relationship between signal propagation and circuit topology, we propose a machine-learning method to predict the circuit timing using complex network features and construct basic and augmented models for comparison. Firstly, in the placement stage, the fundamental circuit features related to timing optimization are extracted, and the complex network features on the timing path are computed by complex network modeling, based on which the dataset for machine learning is generated. Secondly, we use machine learning algorithms to compare the performance of the two models in predicting circuit timing. Experimental results show that the augmented model combining complex network features outperforms the basic model in timing prediction accuracy. The augmented model achieves a 12.5 % reduction in MSE (mean squared error) for predicting slack, reaching 0.004515. It also shows a 1.06 % improvement in , reaching 0.9314, and a 0.54 % improvement in , reaching 0.9652. For predicting delay, the MSE is reduced by 11.69 %, reaching 0.000302, while improves by 0.03 % to 0.9973, and improves by 0.02 % to 0.9987. The deviation in predicting WNS (worst negative slack) is reduced by 6.53 %, reaching 5.30 %, and for TNS (total negative slack), it is reduced by 1.59 %, reaching 12.38 %. The AUC (area under the ROC (receiver operating characteristic) curve) value for identifying critical paths increases by 0.26 %, reaching 0.9873. These results demonstrate the efficiency of the proposed method in utilizing complex network features for predicting circuit timing.
Keywords: Static timing analysis; Complex network; Machine learning; Layout; Routing