Integrated circuit device neural network modeling sample selecting method and device

Abstract

The invention discloses an integrated circuit device neural network modeling sample selecting method and device. With respect to an input variable with a biggest mean impact value, by continually equally diving an interval of the input variable, equal division actions stop until relative errors of all divided intervals are equal to or less than preset error precision, and the length of the divided interval with a smallest length is the step length of the output variable, then step length of each of other input variable is calculated respectively according to the step length of the input variable, finally, with respect to each input variable, points are got according to a change interval and a step length to obtain sample points assembly of each input variable, thus selection of low sample data amount can be achieved under the condition of given precision, and the low sample data amount further saves testing expenditures needed by element molding and increase the training speed of the neural network.

Description

Technical field [0001] This application relates to the field of modeling of integrated circuit devices, and in particular to a method and device for selecting samples for neural network modeling of integrated circuit devices. Background technique [0002] With the development of integrated circuit technology, the feature size of integrated circuit devices continues to decrease. Traditional device models are based on semi-empirical and semi-device physics models. Many simplifications and assumptions have been made for actual devices. This modeling method has been practiced in the past. Medium is feasible, but when the size of the device continues to decrease, it will bring many new physical effects. At this time, according to the traditional modeling method, the model will become complicated or even the modeling of some new devices will become impossible. . [0003] At present, there is a method of modeling integrated circuits using neural networks. Neural networks are an abstracti...

AI Technical Summary

Problems solved by technology

[0005] Too small training samples will cause inaccurate neural network learning, which will affect the final result accuracy

[0006] Excessively large training samples will first cause overfitting, which makes the extrapolation and interpolation capabilities of the neural network very poor, and cannot produce accurate prediction results outside the training samples; too large training samples will also make The burden of training is increased, making the training process more time-consuming, especially when the neural network structure becomes more complex, the time cost will increase exponentially; too large training samples will also increase the time we use for modeling. testing overhead, increasing the cost of modeling

Images