Integrated circuit direct current analysis method and system based on reinforcement learning double-model structure

Abstract

The invention discloses an integrated circuit direct current analysis method and system based on a reinforcement learning double-model structure. According to the method, a step length control strategy of integrated circuit pseudo transient analysis is formulated by using a double-model reinforcement learning algorithm. The reinforcement learning double models comprise a forward model and a backward model, state variables of integrated circuit simulation serve as model input, the state of a current circuit is judged, and an optimal simulation step length is output. According to the integrated circuit direct current analysis method based on the reinforcement learning double-model structure, output of the step length can be self-adapted through different states of the circuit, and the two models are introduced into a common sample pool to mutually learn respective experience, so that the algorithm converges more quickly. Continuous step length output replaces discrete step length output of a traditional algorithm, simulation efficiency can be improved more quickly, and the number of iterations and simulation time of the Newton-Raphson method are greatly reduced.

Description

technical field [0001] The invention relates to integrated circuit simulation technology, in particular to an integrated circuit DC analysis method and system based on a double-model structure of reinforcement learning. Background technique [0002] In the transistor-level circuit simulation of integrated circuit design, calculating the DC operating point is one of the most important and basic tasks. How to successfully obtain a suitable DC operating point has always been the focus of attention in academia and industry. The traditional methods for finding the DC operating point include purePTA, CEPTA, DPTA, etc. The above methods have been studied for several years to calculate the DC operating point through numerical integration. These methods have made important contributions to finding the DC operating point. contribution. However, there are few studies on finding more suitable numerical integration discrete-time control algorithms, and these algorithms have some simulat...

AI Technical Summary

Problems solved by technology

Reinforcement learning algorithms have become more and more mature, but they have not been widely used in the field of transistor-level circuit simulation in integrated circuit design. In this field, the traditional pseudo-transient analysis simulation method uses a fixed discrete step-length multiple to control the next step of the simulation. Step by step, without theoretical analysis and model adaptation for the simulation circuit, it is difficult to solve the problem of simulation efficiency

Images