Neural network for generating composition and parameters of metal alloys based on a given range of known and unknown parameters

Northern Arctic Federal University, Arkhangelsk, 163002, Russia

^* Corresponding author: d.zhuro@narfu.ru

Published online: 10 March 2025

Abstract

This paper describes the development of a generative-adversarial neural network for generating metal alloy compounds with given parameters. The resulting alloy is described by 19 parameters: 14 describe the alloy composition and 5 describe the alloy properties. At the stage of data preparation the parameters are normalized to the range from 0 to 1. The generator in the generative-adversarial network has 4 input layers. The first input layer receives noise to generate different realistic parameters for the same input values. The second input layer is a mask describing the known and unknown parameters. To the third input layer, the minimum acceptable parameter values are passed. To the fourth input layer of the generator the maximum allowable values of parameters are transferred. Based on the input parameters, at the output of the generator we get 19 parameters describing the alloy. The result of the generator is checked by the discriminator for the reliability of the prediction. The discriminator has 4 input layers. The first one receives the prediction made by the generator. The other 3 inputs receive data from the 2^nd, 3^rd and 4^th input layers of the generator. The generative-adversarial neural network is capable of generating the composition and properties of alloys with an average absolute error of 0.082 units relative to the normalized range of test data parameters, i.e. with an accuracy of 91.8% relative to the real value.