Research on thermal management of 3D-ICs assisted by deep learning.

Compared with integrated circuits based on through silicon via (TSV), monolithic inter-tier via (MIV) has been identified as a critical technique to enable three dimensional (3D) integration due to its ultra-small size and relatively superior electrical performance, which allows ultra-high integration density. However, the interconnection of monolithic 3D (M3D) design is more prone to electromigration and stress migration. Severe crosstalk during signal transmission and thermal stress at high temperatures have serious limitations on system performance. In this paper, we focus on the COMSOL Multi-physics software, which can solve multi-field problems, to study the crosstalk problem and its thermal stress problem in MIV structures and analyze the crosstalk effects and temperature stress changes of MIV under different physical coupling conditions. An MIV array based on electrical-thermal-mechanical multi-field coupling was proposed, and the temperature and stress were analyzed by finite element analysis software. Additionally, an artificial neural network scheme is proposed that uses MATLAB to train temperature and stress data to predict the stress values of MIV. Experimental results show that the proposed prediction model using a genetic algorithm to optimize the BP Neural Network (GABP) has a 23.3 % higher prediction accuracy than that of a general BP neural network. • The electrical and thermal behavior characteristics of MIV in M3D were analyzed. • An M3D array is proposed and simulated by FEM. • The data is trained by improved BP network to evaluate the circuit performance. [ABSTRACT FROM AUTHOR]