SIMULATION APPROXIMATORS USING LINEAR AND NONLINEAR INTEGRATED NEURAL NETWORK

Abstract

The construction of approximators for simulations such as the finite element method using machine learning has the problem of a conflict between the reduction of training data generation time and approximation accuracy. Hybrid neural networks have been proposed as a fast approximator for simulations to solve this problem. Training approximators construct it with simple perceptrons with linear activation functions created based on deductive knowledge that can be approximated with less data. However, in the simulation of complex structures, only a limited number of phenomena can be modeled with deductive knowledge. Therefore, there are errors in the predictions. Therefore, a correction approximator has been created by having a neural network learn the errors in the predictions. This approximator is predicted with high accuracy by combining the results of these two approximators. This paper proposes a neural network with a structure that integrates these approximators. Unlike HNN approximators, which prioritize linear approximators, INN optimizes the sharing ratio between linear and nonlinear approximators through learning. This method allows INN to improve the accuracy of approximators and reduce the conflict between the number of training data and accuracy.