Deep Learning Neural Network Algorithm for Computation of SPICE Transient Simulation of Nonlinear Time Dependent Circuits

Abstract

In this paper, a special method based on the neural network is presented, which is conveniently used to precompute the steps of numerical integration. This method approximates the behaviour of the numerical integrator with respect to the local truncation error. In other words, it allows the precomputation of the individual steps in such a way that they do not need to be estimated by an algorithm but can be directly estimated by a neural network. Experimental tests were performed on a series of electrical circuits with different component parameters. The method was tested for two integration methods implemented in the simulation program SPICE (Trapez and Gear). For each type of circuit, a custom network was trained. Experimental simulations showed that for well-defined problems with a sufficiently trained network, the method allows in most cases reducing the total number of iteration steps performed by the algorithm during the simulation computation. Applications of this method, drawbacks, and possible further optimizations are also discussed.

DeepSPICE: Accelerating Digital Cell Characterization Using Deep Learning

“…Recently, there have been research works to speed up this cell characterization process using machine learning methods. These works have addressed either accelerating the circuit simulation time for any given input event combination by speeding up the simulator itself [2], or accelerating the characterization of each cell over a range of manufacturing process variations (the second aspect mentioned in the previous paragraph) [7].…”