CREEM2020

CREEM2020

Prediction of yield and tensile strengths for high-alloy steels from chemical composition: a data preprocessing approach

Submission Author: Giovanni Corsetti Silva , PR
Co-Authors: Giovanni Corsetti Silva, Diogo B. Pitz
Presenter: Giovanni Corsetti Silva

doi://10.26678/ABCM.CREEM2020.CRE2020-0002

 

Abstract

This paper presents the prediction of both yield and tensile strengths for high-alloy steels from chemical composition only, that is, the microstructure, grain size, temperature and other variables are unknown. For that purpose, a neural network was designed, where the input features were: i) raw data and; ii) Regression F-test filtered data. For both cases, a Bayesian optimizer was utilized for tuning. The results show that for the yield strength, preprocessing the data through an F-test gives the best results (R-squared equal to 0.84), meanwhile for the tensile strength, the raw data produces the best performance (R-squared equal to 0.85). Depending on the target and dataset size, removing features through an F-test can substantially increase the accuracy, however, in the scenario where important features are removed, the accuracy dwindles. It was also observed that a convenient way to choose the most efficient neural network is selecting the one with the lowest number of neurons, indicating low overfitting.

Keywords

High-Alloy Steels, Yield Strength, Tensile strength, F-Test, Neural Network

 

DOWNLOAD PDF

 

‹ voltar para anais de eventos ABCM